WO2017210174A1 - Entraînement d'un réseau neuronal à l'aide d'un autre réseau neuronal - Google Patents

Entraînement d'un réseau neuronal à l'aide d'un autre réseau neuronal Download PDF

Info

Publication number
WO2017210174A1
WO2017210174A1 PCT/US2017/034972 US2017034972W WO2017210174A1 WO 2017210174 A1 WO2017210174 A1 WO 2017210174A1 US 2017034972 W US2017034972 W US 2017034972W WO 2017210174 A1 WO2017210174 A1 WO 2017210174A1
Authority
WO
WIPO (PCT)
Prior art keywords
dcnn
transformation
layer
output
document
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/US2017/034972
Other languages
English (en)
Inventor
Uri MERHAV
Dan SHACHAM
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
LinkedIn Corp
Original Assignee
LinkedIn Corp
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Priority claimed from US15/169,063 external-priority patent/US10380480B2/en
Priority claimed from US15/168,750 external-priority patent/US11188823B2/en
Application filed by LinkedIn Corp filed Critical LinkedIn Corp
Priority to CN201780033295.4A priority Critical patent/CN109564636B/zh
Publication of WO2017210174A1 publication Critical patent/WO2017210174A1/fr
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/09—Supervised learning

Definitions

  • the present disclosure generally relates to computer technology for solving technical challenges in machine learning. More specifically, the present disclosure relates to training a neural network using another neural network.
  • a photo may indeed be perfectly shot, w ith perfect lighting and shading, but the subject of the photo may be wearing a stained t-shirt instead of a suit and tie, or may be in a suit and tie but failing to look into the camera.
  • FIG. 1 is a block diagram illustrating a client-server system, in accordance with an example embodiment.
  • FIG. 2 is a block diagram showing the functional components of a social networking service, including a data processing module referred to herein as a search engine, for use in generating and providing search results for a search query, consistent with some embodiments of the present disclosure.
  • a search engine for use in generating and providing search results for a search query, consistent with some embodiments of the present disclosure.
  • FIG. 3 is a block diagram illustrating the application server module of FIG. 2 in more detail.
  • FIG. 4 is a diagram illustrating a Deep Convolutional Neural Network (DCNN), in accordance with an example embodiment.
  • FIG. 5 is a flow diagram illustrating a method of training a DCNN, in accordance with an example embodiment.
  • FIG. 6 is a flow diagram illustrating a method of assigning a professionalism score to an image, in accordance with an example embodiment.
  • FIG. 7 is a block diagram illustrating an image transformation component in greater detail.
  • FIG. 8 is a diagram illustrating an example original image that has been scored for professionalism by the DCNN, in accordance with an example embodiment.
  • FIG. 9 is a diagram illustrating various different possible croppings attempted by the cropping transformation component, as instructed by the optimization problem, solving component.
  • FIG. 10 is a diagram illustrating a selected optimal cropping, in accordance with an example embodiment.
  • FIG. 11 is a diagram illustrating various different possible rotations attempted by the rotation transformation component, as instructed by the optimization problem solving component.
  • FIG. 12 is a diagram illustrating a selected optimal rotation in accordance with an example embodiment.
  • FIG. 13 is a flow diagram illustrating a method for automatically cropping a digital image to improve perceived professionalism of a subject of the image, in accordance with an example embodiment.
  • FIG. 14 is a flow diagram illustrating a method for automatically performing an image transformation on a digital image to improve perceived professionalism of a subject of the image, in accordance with an example embodiment.
  • FIG. 15 is a flow diagram illustrating a method for filtering out outliers in a Stochastic Grandient Descent optimization in a DCNN, in accordance with an example embodiment.
  • FIG. 16 is a flow diagram illustrating a method for training a DCNN, in accordance with an example embodiment.
  • FIG. 17 is a block diagram illustrating the application service module after a switchover to a second DCNN has occurred, in accordance with an example embodiment.
  • FIG. 18 is a flow diagram illustrating a method of training and utilizing DCN s, in accordance with an example embodiment.
  • FIG. 19 is a flow diagram illustrating a method for changing a system over from a first DCNN to a second DCNN, in accordance with an example embodiment.
  • FIG. 20 is a block diagram illustrating a representative software architecture, which may be used in conjunction with various hardware architectures herein described.
  • FIG. 21 is a block diagram illustrating components of a machine, according to some example embodiments, able to read instructions from a machine-readable medium (e.g., a machine-readable storage medium) and perform any one or more of the methodologies discussed herein.
  • a machine-readable medium e.g., a machine-readable storage medium
  • a DCNN is used to generate
  • the DCNN uses information about digital images previously categorized with professionalism scores or labels indicating a level of professionalism to train the DCNN to identify relevant features of input images. Relevant features may include any features that the
  • the DCNN deems important to a determination of the professionalism (e.g., useful in calculating a professionalism score) of an input image.
  • the DCNN is able to self-identify these relevant features, without the need for a human to predetermine which features may be important and instruct a machine learning algorithm to extract and examine such features, as is necessary in more traditional machine learning techniques.
  • the DCNN is free to identify any feature that tends to be shared among images of high professionalism, regardless of whether one would ordinarily think of such a feature as one that might pertain to professionalism.
  • image lighting, clothing of subject matter, and angle of head may be features that one might expect to affect the perceived professionalism of a subject in an image, features such as how the image is cropped, color balance, and depth of field might not.
  • the DCNN is able to itself determine that these features, even if they are subtle and perhaps not even detectable by a human, are important to a perception of professionalism and adjust its model accordingly.
  • FIG. 1 is a block diagram illustrating a client-server system 100, in accordance with an example embodiment.
  • a networked system 102 provides server-side functionality via a network 104 (e.g., the Internet or a wide area network (WAN)) to one or more clients.
  • FIG. 1 illustrates, for example, a web client 106 (e.g., a browser) and a programmatic client 108 executing on respective client machines 110 and 112.
  • a web client 106 e.g., a browser
  • programmatic client 108 executing on respective client machines 110 and 112.
  • An application program, interface (API) server 114 and a web server 1 16 are coupled to, and provide programmatic and web interfaces respectively to, one or more application servers 118.
  • the application server(s) 118 host one or more applications 120.
  • the application server(s) 118 are, in turn, shown to be coupled to one or more database servers 124 that facilitate access to one or more databases 126. While the application(s) 120 are shown in FIG. 1 to form part of the networked system 102, it will be appreciated that, in alternative
  • the application(s) 120 may form, part of a service that is separate and distinct from the networked system 102.
  • client-server system 100 shown in FIG. 1 employs a client-server architecture
  • present disclosure is, of course, not limited to such an architecture, and could equally well find application in a distributed, or peer-to-peer, architecture system, for example.
  • the various applications 120 could also be implemented as standalone software programs, which do not necessarily have networking capabilities.
  • the web client 106 accesses the various applications 120 via the web interface supported by the web server 116.
  • the programmatic client 108 accesses the various services and functions provided by the application ⁇ ⁇ 120 via the programmatic interface provided by the API server 1 14.
  • FIG. 1 also illustrates a third party application 128, executing on a third party server 130, as having programmatic access to the networked system 102 via the programmatic interface provided by the API server 114.
  • the third party application 128 may, utilizing information retrieved from the networked system 102, support one or more features or functions on a website hosted by a third party.
  • the third party website may, for example, provide one or more functions that are supported by the relevant applications 120 of the networked system 102.
  • any website referred to herein may comprise online content that may be rendered on a variety of devices including, but not limited to, a desktop personal computer (PC), a laptop, and a mobile device (e.g., a tablet computer, smartphone, etc.).
  • a user can use a mobile app on a mobile device (any of the machines 110, 112 and the third party server 130 may be a mobile device) to access and browse online content, such as any of the online content disclosed herein.
  • a mobile server e.g., API server 114 may communicate with the mobile app and the application server(s) 118 in order to make the features of the present disclosure available on the mobile device.
  • the networked system 102 may comprise functional components of a social networking service.
  • FIG. 2 is a block diagram showing the functional components of a social networking sendee, including a data processing module referred to herein as a search engine 216, for use in generating and providing search results for a search query, consistent with some embodiments of the present disclosure.
  • the search engine 216 may reside on the application server(s) 1 18 in FIG. 1. However, it is contemplated that other configurations are also within the scope of the present disclosure.
  • a front end may comprise a user interface module (e.g., a web server 116) 212, which receives requests from various client computing devices, and communicates appropriate responses to the requesting client devices.
  • the user interface module(s) 212 may receive requests in the form of Hypertext Transfer Protocol (HTTP) requests or other web-based API requests.
  • HTTP Hypertext Transfer Protocol
  • a member interaction detection module 213 may be provided to detect various interactions that members have with different applications 120, services, and content presented. As shown in FIG. 2, upon detecting a particular interaction, the member interaction detection module 213 logs the interaction, including the type of interaction and any metadata relating to the interaction, in a member activity and behavior database 222.
  • An application logic layer may include one or more various application server modules 214, which, in conjunction with the user interface module(s) 212, generate various user interfaces (e.g., web pages) with data retrieved from various data sources in a data layer.
  • individual application server modules 214 are used to implement the functionality associated with various applications 120 and/or sen/ices provided by the social networking service.
  • the data layer may include several databases, such as a profile database 218 for storing profile data, including both member profile data and profile data for various organizations (e.g., companies, schools, etc.).
  • a profile database 218 for storing profile data, including both member profile data and profile data for various organizations (e.g., companies, schools, etc.).
  • the person when a person initially registers to become a member of the social networking service, the person will be prompted to provide some personal information, such as his or her name, age (e.g., birthdate), gender, interests, contact information, home town, address, spouse's and/or family members' names, educational background (e.g., schools, majors, matriculation and/or graduation dates, etc.), employment history, skills, professional organizations, and so on.
  • This information is stored, for example, in the profile database 218.
  • the representative may be prompted to provide certain information about the organization.
  • This information may be stored, for example, in the profile database 218, or another database (not shown).
  • the profile data may be processed (e.g., in the background or offline) to generate various derived profile data. For example, if a member has provided information about various job titles that the member has held with the same organization or different organizations, and for how long, this information can be used to infer or derive a member profile attribute indicating the member's overall seniority level, or seniority level within a particular organization.
  • importing or otherwise accessing data from one or more externally hosted data sources may enrich profile data for both members and organizations. For instance, with organizations in particular, financial data may be imported from one or more external data sources and made part of an organization's profile. This importation of organization data and enrichment of the data will be described in more detail later in this document.
  • a member may invite other members, or be invited by other members, to connect via the social networking service.
  • a "connection” may constitute a bilateral agreement by the members, such that both members acknowledge the establishment of the connection.
  • a member may elect to "follow" another member.
  • the concept of " 'following' 1 another member typically is a unilateral operation and, at least in some embodiments, does not require acknowledgement or approval by the member that is being followed.
  • the member who is following may recei e status updates (e.g., in an activity or content stream) or other messages published by the member being followed, or relating to various activities undertaken by the member being followed.
  • the member when a member follows an organization, the member becomes eligible to receive messages or status updates published on behalf of the organization. For instance, messages or status updates published on behalf of an organization that a member is following will appear in the member's personalized data feed, commonly referred to as an activity stream or content stream.
  • the various associations and relationships that the members establish with other members, or with other entities and objects, are stored and maintained within a social graph in a social graph database 220.
  • the members' interactions and behavior e.g., content viewed, links or buttons selected, messages responded to, etc.
  • the members' interactions and behavior may be tracked, and information concerning the members " activities and behavior may be logged or stored, for example, as indicated in FIG. 2, by the member activity and behavior database 222.
  • This logged activity information may then be used by the search engine 216 to determine search results for a search query.
  • the databases 218, 220, and 222 may be incorporated into the database(s) 126 in FIG. 1.
  • other configurations are also within the scope of the present disclosure.
  • the social networking service system 210 provides an API module via which applications 120 and services can access various data and services provided or maintained by the social networking service.
  • an application may be able to request and/or receive one or more navigation recommendations.
  • Such applications 120 may be browser-based applications 120, or may be operating system-specific.
  • some applications 120 may reside and execute (at least partially) on one or more mobile devices (e.g., phone or tablet computing devices) with a mobile operating system.
  • the applications 120 may be applications 120 and sen/ices that are developed and maintained by the entity operating the social networking service, nothing other than data privacy concerns prevents the API from being provided to the public or to certain third parties under special arrangements, thereby making the navigation recommendations available to third party applications 128 and services.
  • search engine 216 is referred to herein as being used in the context of a social networking service, it is contemplated that it may also be employed in the context of any website or online services. Additionally, although features of the present disclosure are referred to herein as being used or presented in the context of a web page, it is contemplated that any user interface view (e.g., a user interface on a mobile device or on desktop software) is within the scope of the present disclosure.
  • search engine 216 facilitates the indexing and searching for content within the social networking service, such as the indexing and searching for data or information contained in the data layer, such as profile data (stored, e.g., in the profile database 218), social graph data
  • member activity and behavior database 222 stored, e.g., in the social graph database 220
  • member activity and behavior database 222 stored, e.g., in the member activity and behavior database 2202.
  • the search engine 216 may collect, parse, and/or store data in an index or other similar structure to facilitate the identification and retrieval of information in response to received queries for information. This may include, but is not limited to, forward search indexes, inverted indexes, N-grarn indexes, and so on.
  • FIG. 3 is a block diagram illustrating application server module 214 of FIG. 2 in more detail. While in many embodiments the application server module 214 will contain many subcomponents used to perform various different actions within the social networking system, only those components that are relevant to the present disclosure are depicted in FIG. 3.
  • An image ingestion component 300 obtains images from profile database 218. These images may be, for example, images uploaded by members to their own member profiles. In some example embodiments, the image ingestion component 300 may obtain the images directly from the member via a user interface, rather than the images being first stored in the profile database 218.
  • An image preprocessing component 302 takes each image from the image ingestion component 300 and preprocesses it. This may include, as will be described in more detail below, normalizing certain aspects (such as aspect ratio, length, width, color channels, etc.) of the image. Each preprocessed image is then passed to a first DCNN 304, which has been trained using training images to analyze the preprocessed image and produce a professionalism score that indicates a level of professionalism in the content of the image. This process will be described in more detail below. This professionalism score will then be used by an image transformation component 306 to transform some aspect of the image, such as cropping, rotation, lighting, shading, sharpness, and the like based on the professionalism score to improve the professionalism of the subject of the image. This will also be described in more detail below. The transformed image may then be stored in the profile database 218.
  • the first DCNN 304 is a machine-learning model that effectively infers non-linear relationships between a homogeneous input field and desired outputs, which are either categorical classes or scaiars.
  • the first DCNN 304 is a model that maps inputs to outputs using a sequence of so-called convolutionai layers of artificial neurons.
  • the first DCNN 304 may be trained by presenting it with a large number (e.g., greater than 10,000) of sample images and labels (e.g.,
  • the model may be applied to new input images to produce a useful prediction of the professionalism levels of the new input images.
  • the first DCNN 304 performs better when the input images have been normalized, and so a preprocessing operation may be performed by the image preprocessing component 302 on both the sample images and the input images to normalize them. This normalization may include, for example, altering all images to be of a consistent size (e.g., 250 x 250 pixels) and of consistent color channels (e.g., 3 color channels of red-green-blue).
  • the first DCNN 304 also performs better when the inputs are homogenous, meaning that the inputs have a uniform meaning (e.g., each pixel represents intensity), and neighborhood relationships in the input are well defined (e.g., pixels adjacent to each other represent physically close locations).
  • the first DCNN 304 is designed to learn not only
  • classifications/professionalism scores for images but also to leam the feature hierarchy, all the way from pixels to classifier, by defining a number of layers.
  • the process of inference involves taking a given input, applying a sequence of mathematical functions called layers, and calculating the functions on the input images. Each layer extracts features from the output of a previous layer, and all layers are trained jointly.
  • the layer-based architecture is why it is termed a "deep" convolutional neural network.
  • FIG. 4 is a diagram illustrating a first DCNN 400, in accordance with an example embodiment.
  • the first DCNN 400 may be, in some example embodiments, the DCNN 304 of FIG. 3. Here, two stages 402A, 402B are depicted.
  • the convolutional layers 404A, 404B are the core of the first DCNN 400. Their parameters include a set of learnable filters that have a small receptive field, but extend through the full depth of the input image. In other words, for color images, each filter may look at a relatively small number of pixels in each image but look at all the colors in that area.
  • each filter is convolved across the width and height of the input image, computing the dot product between the entries of the filter and the input and producing a two-dimensional activation map of that filter.
  • the first DCNN 400 learns filters that activate when they see some specific type of feature at some spatial position in the input image.
  • the feature maps for all filters can be stacked along the depth dimension to form the full volume output of the convoiutionai layers 404A, 404B.
  • the convoiutionai layers 404A, 404B apply mathematical operations called convoiutionai.
  • the filters used in the convoiutionai layers 404A, 404B may be activated in a first iteration of the first DCNN 400 and refined prior to each additional iteration, based on actions taken in other layers in the previous iteration, until some error term is minimized below a particular threshold. In one example embodiment, this may be accomplished through back propagation, which is described in more detail below.
  • each feature map 406A-406C may be produced by a different filter and modified based on various functions in each stage. For example, if the input is a color image, each feature map 406A-406C may be a two-dimensional array of some aggregation of all color channel information. At the output, each feature map 406A-406C represents a particular feature extracted at all locations on the input and conditioned. The example in FIG.
  • the nonlineanty layers 408A, 408B give the first DCNN 400 greater expressive power in uncovering nonlinear relationships between input and output.
  • Many different nonlinearities could be used in the nonlinearity layer, including sigmoid, tanh, and rectified linear function.
  • the rectified linear function For brevity, one example of nonlinearity will be described here: the rectified linear function. This function is defined by the followin
  • the pooling layers 41 OA, 410B are applied to lower the input image's spatial dimensions while preserving some information from the input image.
  • the pooling layers 41 OA, 410B do not actually do any of the learning; in other words, they are a fixed predefined operation that does not change as training progresses. Instead, they are used the spatial dimensions of the problem.
  • a decimation approach could be followed, where one out of eve ' N samples along a spatial dimension is kept out.
  • some local statistics may be used for pooling, such as max pooling, defined as:
  • a classification layer 412 is used to classify the image using the output of the final pooling layer 410B.
  • the classification layer 412 is actually a specialized convolutional layer containing a filter designed to produce the classification score from the volume output of the final pooling layer 410B.
  • This filter applies a classification function having weights that may be refined in the same manner as the weights in the fun ctions of the filters of the norm al con volutional layers 404, 404B.
  • Back propagation involves calculating a gradient of a loss function (defined later) in a loss layer 414, with respect to a number of weights in the first DCNN 400.
  • the gradient is then fed to a method that updates the weights for the next iteration of the training of the first DCNN 400 in an attempt to minimize the loss function, which uses a different plurality of sample images (unless there is a need to repeat, such as running out of sample images).
  • Back propagation uses the labeled sample images in a batch of sample images that have been passed through the stages 402A, 402B in order to calculate the loss function gradient for the samples as a group (although, as will be seen later, the loss function may be modified dynamically to eliminate some of the samples from consideration).
  • Back propagation may include two aspects: propagation and weight update.
  • forward propagation of a training pattern's input images is performed through the first DCNN 400 in order to generate the propagation's output activations (i.e., the images are passed through the stages 402A, 402B).
  • backward propagation of the propagation's output activations are performed through the DCNN 400 using a target specified by the training pattern in order to generate the deltas of ail output.
  • the output delta and input activation are multiplied to obtain the gradient of the weight, and then a ratio of the gradient is subtracted from the weight.
  • the ratio influences speed and quality of learning. The higher the ratio, the faster the training, but at the expense of accuracy.
  • the use of the back propagation may be predicated on whether or not the combined error of the classification of the images in the batch of labeled sample images transgressed a preset error threshold. If the combined error is too great, then back propagation should occur to update and hopefully minimize the error for the next iteration, and a next, iteration is performed with a subsequent batch of labeled sample images, until the combined error does not transgress the threshold.
  • the labeled output may be a professionalism score for the image, or may be a categorical variable (e.g., high professionalitv, medium professionalitv, or low professionalitv).
  • the first DCNN 400 outputs a vector that may be compared to the desired output of some loss function, such as the sum square error function:
  • the DCNN is trained to output a vector which represents the probability of evesy class, and some probabilistic loss function, such as the known entry cost function, may be used as follows:
  • the output from the first DCNN 304 may be passed to an image transformation component 306 that acts to transform a candidate image based on the professionalism score of the image (as well as potentially professionalism scores of various transformed versions of the image). This process will be described in more detail below.
  • the transformed image from the image transformation component 3 6 can be used as a profile image and thus stored in profile database 218, as well as being used to train a second DCNN 308.
  • the second DCNN 308 may be constructed in a similar fashion as the first DCNN 304, except that it is trained to output a single " 'optimal" transformation for an input image, as opposed to outputting a professionalism score for an output. The exact manner of this output may differ based on the type of transformation for which the second DCNN 308 is trained.
  • the training data may comprise sample output croppings from image transformation component 306 as performed on candidate images based on the professionalism scores from the first DCNN 304, but the second DCNN 308 may output
  • FIG. 5 is a flow diagram illustrating a method 500 of training a first DCNN 400, in accordance with an example embodiment.
  • a batch of sample labeled images are fed to the first DCNN 400 and the current model of the first DCNN 400 produces an output.
  • This output may be, for example, a classification of professionalism or a professionalism score for each sample labeled image.
  • the loss layer 414 of the first DCNN 400 calculates the error for the batch of sample labeled images. This error may be, for example, a combination of the individual errors for each of the individual sample labeled images.
  • weights inside the filters in the convolutional layers 404A, 404B are updated to minimize the loss, in accordance with the loss function defined in the loss layer 414.
  • the defined validation set is based on a stochastic gradient descent (SGD) function.
  • SGD comprises the following steps:
  • FIG. 6 is a flow diagram illustrating a method 600 of assigning a professionalism score to an image, in accordance with an example embodiment.
  • the professionalism refers to a perceived professionalism of the subject of the image, with the subject being a human (e.g., how professional the human appears in his or her job), as opposed to, for example, the professionalism of how the subject was photographed.
  • the image is passed to a convo!utiona! layer of a DCNN, generating output.
  • the output from the convolutional layer is passed to a nonlinearity layer, generating output.
  • the output from the nonlinearity layer is passed to a pooling layer, generating output.
  • the output from the nonlinearity layer is passed to a classification layer, generating a professionalism score for the digital image. It should be noted that in tins runtime process where a professionalism score is assigned to an unlabeled image, there is no need to utilize a loss layer (indeed, use of the loss layer would not make sense as there is no label for the image to be used to determine the error level).
  • the professionalism score generated by the DCNN 304 for an image may be used by the image transformation component 306 in performing one or more transformations on the image.
  • transformations include, but are not limited to, cropping, rotating, lighting compensation, edge enhancing filtering, blur filtering, and spatially varying blur filtering.
  • FIG. 7 is a block diagram illustrating an image transformation component 306 in greater detail.
  • An optimization problem solving component
  • An optimization problem is the problem of finding a best solution from all feasible solutions.
  • the optimization problem may be one with discrete variables, and thus is known as a combinatorial optimization problem.. This is because the optimization problem attempts to find the best transformation among all the possible transformations of the input image.
  • the optimization problem solving component 700 attempts to solve the optimization problem using an optimization algorithm .
  • optimization algorithms for discrete variables include deterministic optimization algorithms such as Gauss-Newton, Levenberg- Marquardt, Quasi-Newton, Nonlinear Conjugate Gradient, and Step-size x, as well as stocliastic algorithms such as SGD, Robbin-Monro, and decaying sequence.
  • a continuous optimization problem algorithm may be utilized.
  • optimization algorithms for continuous variables include unconstrained optimization algorithms, such as derivative-free algorithms, line-search methods, gradient methods, Newton and trust region methods, and algorithms based on conjugate directions, as well as constrained optimization algorithms, such as linear optimization (pivoting algorithms and interior point methods), convex quadratic optimization, and general nonlinear optimization algorithms (duality theory, reduced gradient methods, and barrier methods).
  • the optimization problem solving component 700 will interact with the DCNN 304 for feedback as to the effect of various different possible transformations on an image, which are defined by the various individual transformation components 702A-702F.
  • These individual transformation components 702A-702F each apply a different type of transformation to the image, using parameters specified by the optimization problem solving component 700. In that manner, the individual transformation components 702A-702F do not need to each run every possible transformation of the appropriate type on the image. For example, it is not necessary for the cropping transformation component 702 A to perform all the possible different ways to crop a particular image so that each of those possible crops can be assigned a professionalism score by the DCNN 304 so that the optimization problem solving component 700 can select the cropping that increases the professionalism score the most, but rather the optimization problem solving component 700 can select certain particular croppings having the highest likelihood of success and request only those croppings be investigated further by specifying them in parameters to the cropping transformation component 702A.
  • FIG. 8 is a diagram illustrating an example original image 800 that has been scored for professionalism by die DCNN 304, in accordance with an example embodiment.
  • the optimization problem solving component 700 may undertake to determine the optimal crop of this original image 800 so as to maximize the
  • FIG. 9 is a diagram illustrating various different possible croppings attempted by the cropping transformation component 702A, as instructed by the optimization problem solving component 700. Each of these different possible croppings may be scored by the DCNN 304, and the optimization problem solving component 700 selects the cropping that increases the professionalism score the most.
  • FIG. 10 is a diagram illustrating a selected optimal cropping 1000, in accordance with an example embodiment. It should be noted that while in some instances the result of the optimization problem may be the selection of an optimal cropping that is similar to what would be selected as an optimal cropping if a different parameter were being examined, such as quality of the image itself, in many cases the professionalism score will result in a completely different selection than these other types of parameters. For example, the optimal cropping for professionalism may be a tighter head shot than the optimal cropping for image quality, which might be more interested in centering the subject in the shot than on the amount of or lack of background imagery in the shot.
  • the automatic cropping process described above may be improved by training a separate neural net that sees a picture as an input and outputs the best rectangle from the picture. In this manner, it is not necessary to have a human annotate images for the "best rectangle.” Instead, the automatic cropping process is used to find a crop that maximizes the score, and that crop is used as the "truth" that the other neural net is training for. In a sense, the output of one DCNN is used as a label for input to tram a different neural net.
  • the second neural net uses this technique to express learning rales such as "usually the object of interest is around the middle of the image” or "the object of interest should never have a very small width or height.” This improves performance over a process that scores rectangles without spatial context.
  • the rotation transformation component 702B acts to perform rotation transformations on input images, in accordance with parameters specified by the optimization problem solving component 700.
  • FIG. 11 is a diagram illustrating various different possible rotations attempted by the rotation transformation component 702B, as instructed by the optimization problem solving component 700. Each of these different possible rotations may be scored by the DCNN 304, and the optimization problem solving component 700 selects the rotation that increases the professionalism score the most.
  • FIG. 12 is a diagram illustrating a selected optimal rotation 1200, in accordance with an example embodiment.
  • the lighting compensation transformation component 702C acts to perform lighting compensation transformations on input images, in accordance with parameters specified by the optimization problem solving component 700. These lighting compensation transformations may include lightening or darkening areas of the image.
  • Hie edge enhancing filtering transformation component 702D acts to perform edge enhancement filtering on input images, in accordance with parameters specified by the optimization problem solving component 700. Edge enhancement filtering enhances the edge contrast of an image in an attempt to improve its sharpness. The filter works by identifying sharp edge boundaries in the image, such as the edge between a subject and the background, and increasing the image contrast in the area immediately around the edge.
  • the blur filtering transformation component 702E acts to perform blur filtering transformations on input images, in accordance with parameters specified by the optimization problem solving component 700.
  • the blur filtering transformations are designed to smooth out skin texture.
  • the spatially varying blur filtering transformation component 702F acts to perform spatially varying blur filtering tra sformations on input images, in accordance with parameters specified by the optimization problem solving component 700.
  • spatially varying blur filtering involves blurring out the background, but not the subject.
  • the optimization problem solving component 700 can interact with each of the individual transformation components 702A-702F individually, or alternatively may interact with groupings of the individual transformation components 702A-702F simultaneously. For example, if the optimization problem solving component 700 wishes to determine the best combination of cropping and rotation that can be performed on the image in order to improve the professionalism the most, the optimization problem solving component 700 can interact with both the cropping transformation component 702A and the rotation transformation component 702B simultaneously to arrive at a joint result of the best cropping/rotation combination, or may interact with each individually, producing an independent answer for best cropping and for best rotation.
  • the aspects described above with respect to the image transformation component 306 need not necessarily operate in conjunction with a first DCNN 304.
  • a different machine learning component may be utilized to generate professionalism scores used by the image tra sformation component 306 in determining the best transformation to perform on an image.
  • FIG. 13 is a flow diagram illustrating a method 1300 for
  • a machine learning algorithm is used to generate a
  • this machine learning algorithm may be the
  • the professionalism score is used as an input to a discrete variable optimization algorithm to determine an optimal cropped version of the digital image from a plurality of possible cropped versions of the digital image using the classification function.
  • FIG. 14 is a flow diagram illustrating a method 1400 for
  • a machine learning algorithm is used to generate a professionalism score for the digital image, utilizing a machine learning algorithm, comprising: a training mode where a plurality of sample images with labeled professionalism scores are used to train a classification function in a model that produces a professionalism score as output; and an analysis mode where the model is used to generate a professionalism score for the digital image.
  • this machine learning algorithm may be the DCNN 304 or 400 described above, although this is not mandatory.
  • the professionalism score is used as an input to a continuous variable optimization algori thm, to determine an optimal version of the digital image from a plurality of possible versions of the digital image on which one or more image transformations have been performed, using the classification function.
  • the professionalism scores produced by the model may be used to identify the best image from a group of images for a member to use as a profile picture.
  • a system may be designed to automatically browse through images on the user's device and select an image showing the member, with the image selected being the one with the highest professionalism score.
  • this may be performed by limiting the image search to images labeled "seifie” such as by being stored in a directory with the name "seifie" in it.
  • the selfies themselves may be identified using a machine learning algorithm.
  • social network information from, for example, the profile database 218 may be used in conjunction with the first DCNN 304 described above with respect to the professionalism of images posted or uploaded to the social network sendee.
  • Social networks often have very abundant information that can be used to aid in the training of the first DCNN 304, as not only image information is available but also various pieces of information about the subject of the images is also available, such as job title, experience level, skills, age, and so forth.
  • an automatic labelling component may be added to the system described above in FIG. 3 and the corresponding text to apply automatic labels to sample images based on various signals taken from a social networking profile associated with each sample image.
  • the automatic labelling component may parse fields on the social networking profile to deduce a level of professionalism for the corresponding member, exclusive of any images. Fields that generally may be used to deduce a level of professionalism include title, skills, experience, and what types of other social networking members are linked to the social networking profile, among others.
  • the automatic labelling component may use its own machine learning algorithm to train a model to assign a level of professionalism to social media profiles to which one or more sample images are linked.
  • the machine learning algorithm may be selected from among many different potential supervised machine learning algorithms. Examples of supenised learning algorithms include Bayesian networks, instance-based learning, support vector machines, random forests, linear classifiers, quadratic classifiers, k-nearest neighbor, decision trees, and hidden Markov models.
  • the automatic labelling component may simply make a recommendation to a classifier as to a label to apply to an image.
  • the classifier may still need to approve the recommendation in order for it to be applied to the image, and the classifier may override the recommendation.
  • the automatic labelling component can use information from the social media profiles to improve labelling, the social media profiles often can be rife with outliers and incorrect information (e.g., members claiming to be 130 years old, merely because they did not want to put in a birthdate). This gives rise to a need to automatically ignore such outliers during training and zoom in on a correct solution in a robust way that does not require human handling to exclude the outliers. It may also be difficult for the automatic labelling component to compensate for these errors, or otherwise be modified to handle these errors.
  • a specialized procedure may be implemented in the loss layer 414 of the first DCNN 400.
  • the loss layer 414 might ordinarily apply a static loss function to its input.
  • the loss layer 414 is modified to include a dynamically changing function that adjusts based on statistical analysis of the samples.
  • FIG. 15 is a flow diagram illustrating a method 1500 for filtering out outliers in a SGD optimization in a DCNN, in accordance with an example embodiment.
  • a batch of N samples fed to the DCNN are randomly selected.
  • the DCNN is utilized to make an assessment of professionalism of each of the N samples.
  • a static loss function may be applied (called loss[n ⁇ ).
  • the calculated losses are sorted and enumerated in tuples of (rank, loss).
  • a modified loss function (k ⁇ rank, loss)) is introduced that is dynamic and applied to the losses, producing
  • the dynamic loss function may be as follows:
  • the dynamic loss function may be more nuanced, applying statistical tests. For example, a Gaussian distribution of errors may be assumed, and the samples weighted by their chances of violating the Gaussian assumption. This means the mean ⁇ , standard deviation ⁇ in the current batch, and the measurements may be normalized as follows:
  • the samples may be weighted by their probability of not belonging to the error statistics:
  • the DCNN may output professionalism scores of ( 1.1, 1.2, 2, 2, and 97 ⁇ for the five samples, but the labels for the images were assigned as (1, 1 , 2, 2, 1 ). If the static loss function is absolute error (prediction-correct), then the losses from the static loss function are (0.1, 0.2, 0, 0, 96). The ranking of the samples is then (3, 4, 1, 2, 5), as 5 was clearly the most imprecise in comparison to its expected value. If the simple dynamic loss function described above, which ignores the top 20% worst samples, is applied, then the dynamic loss function would filter out the result for sample 5 in the above batch, since it is among the 20% with the lowest rank. Thus, the results from the modified dynamic loss function are (0.1, 0,2, 0, 0, 0). As such, the outlying sample does not motivate any changes in the model to accommodate it.
  • the dynamic loss function approach may be extended by calculating the error statistics on multiple recent batches together as opposed to the current batch only.
  • the determination of which samples in a most recent batch to ignore by the dynamic loss function is determined based on statistic information from the last N batches, where N is greater than I .
  • a schedule of stricter outlier thresholds is introduced as the training progresses (such that outliers are not ignored when the model starts learning, and progressively more aggressively ignored as the training process contains).
  • k(loss, rank, batch ) the number of times that outliers are not ignored when the model starts learning, and progressively more aggressively ignored as the training process contains.
  • FIG. 16 is a flow diagram illustrating a method 1600 for training a DCNN, in accordance with an example embodiment.
  • the DCNN in this method 1600 may or may not be the first DCNN 304 or 400, or the second DCNN 308 described above.
  • the method 1600 may utilize a DCNN that is deigned to output a professionalism score for an image, or may utilize a DCNN that is designed to output any other sort of classification on any other type of input.
  • Operation 1604A involves passing the sample to a convolutional layer of the DCNN, with the convolutional layer comprising one or more filters having dynamically adj ustable weights, and the one or more filters configured to filter the sample to produce an output volume for the corresponding sample, with the output volume comprising a different feature map for each of the one or more filters.
  • Operation 1604B involves passing the output volume from the convolutional layer through a nonlinearity layer, with the nonlinearity layer applying a nonlinearity function to the output volume from the convolutional layer.
  • Operation 1604C involves passing the output volume from the nonlinearity layer through a pooling layer, with the pooling layer lowering spatial dimensions of the output volume from the nonlinearity layer.
  • Operation 1604D involves passing the output volume from the pooling layer through a classification layer, with the classification layer comprising a specialized convolutional layer having a filter designed to output a prediction for the sample based on the output volume from the pooling layer.
  • Operation 1604E involves passing the sample through a loss layer, with the loss layer applying a loss function to the sample, resulting in an indication of a level of error in the prediction from the classification layer in comparison to the label of the sample.
  • each of the current plurality of samples is ranked based on its corresponding level of error.
  • a dynamic loss function is applied to the current plurality of samples to eliminate lower ranked samples from consideration .
  • the dynamic loss function may be based on statistics regarding the current plurality of samples, or may be based on statistics regarding a combination of the current plurality of samples and one or more previous plurality of samples used previously to train the DCNN.
  • the dynamic loss function may be designed to become automatically stricter as more iterations of the training occur.
  • weights of the one or more filters in the convolutional layers (including the classification layer) of the DCNN are updated to reduce the combination of the levels of error. 1 hen the method 600 loops back to operation 1602 to repeat the training with a next plurality of samples.
  • the output from the image transformation component 306 may be used to train a second DCNN 308 to perform image transformation without the same computationally intensive processes used by the image transformation component 306.
  • the second DCNN 308 is trained to perform a specific type of transformation (e.g. cropping)
  • only a single second DCNN 308 will be described herein, and it will be described in the context of the cropping operation.
  • Training the second DCNN 308 is beneficial from the technical point of view because the image transformation component 306 may be designed to perform its transformations without the benefit of contextual information about the image. For example, in the case of croppings, the image transformation component 306 essentially attempts multiple different croppings and scores each one. While the croppings selected to be attempted may not exactly be random (see the discussion about the optimization problem solving above), they are not selected based on contextual information about the image. For example, often the image preprocessing component 302, in an effort to standardize the size of images passed to the first DCNN 304, will stretch or shrink one or both dimensions of an image.
  • the image transformation component 306 then does not have information about whether the image it is transforming has been stretched or not, and this information may be useful in determining what is an optimal crop for the image (a cropping that results in a warped picture, for example, will not be optimal). Additionally, there are certain sizes of croppings that simply do not make sense given the size of an image or the subject matter. If we are dealing with profile pictures (and thus, subjects' faces), rectangles that are too short in comparison with their width or two narrow in comparison with their height are likely not to be good selections, nor are rectangles that are simply too small (e.g., 3 pixels x 4 pixels).
  • the DCNN 308 By training the DCN 308 to perform image transformations without image preprocessing, the DCNN 308 is able to learn contextual-based features of not just the potential cropping area but also surrounding areas, making it able to select the optimal crop quicker, and thus without using as many processing cycles.
  • a "switchover" may occur where new candidate images are passed to the second DCNN 308 rather than being passed to the image preprocessing component 302, first DCNN 304, and image transformation component 306.
  • a DCNN selector 310 may make this determination as to whether the second DCNN 308 has been trained enough to make the switchover occur. The decision-making process of the DCNN selector 310 will be described in more detail below.
  • FIG. 17 is a block diagram illustrating application service module 214 after a switchover to a second DCNN 308 has occurred, in accordance with an example embodiment.
  • the second DCNN 308 has been fully trained, and thus candidate images are passed directly from the image ingestion component 300 to the second DCNN 308, bypassing the image preprocessing component 302, first DCNN 304, and image transformation component 306.
  • this figure depicts the image processing component 302, first DCNN 304, and image transformation component 306 as being bypassed, in some example embodiments, this bypassing may only be occurring for the particular type of transformation that the second DCNN 308 has been trained for.
  • cropping on images may occur directly from the image ingestion component 300 using the second DCNN 308, but other types of transformations (e.g., rotations, lighting adjustments, etc.) may continue to be performed by the image transformation component 306 via the image preprocessing component 302 and first DCNN 304, as described above with respect to FIG. 3.
  • other types of transformations e.g., rotations, lighting adjustments, etc.
  • the DCNN selector 310 may determine that it is time to switch over from the first DCNN 304 to the second DCNN 308 by utilizing one or more metrics to evaluate the output of the first DCNN 304 and the second DCNN 308.
  • the metric used is actually the professionalism score output by the first DCNN 304.
  • the image transformation component 306 may already have obtained a professionalism score for the cropping that it felt was the optimal cropping for a particular image, given that the professionalism score is how it determines optimal transformations.
  • the image tra sformation component 306 already has a professionalism score for the optimal cropping produced by the first DCNN 304 route (for simplicity called the " 'first optimal cropping").
  • the DCNN selector 310 can then take, for the same input image, the optimal cropping output by the second DCNN 308 (for simplicity called the "second optimal cropping") and feed that into the first DCNN 304 to produce a professionalism score
  • the DCNN selector 310 determines that it is time to changeover from the first DCNN 304 to the second DCNN 308.
  • the DCNN selector 310 may go a step further. It should be appreciated that, independent of accuracy in selecting the optimal cropping, use of the second DCNN 308 is much less computing and resource intensive than use of the first DCNN 304. As such, it may not be necessary to wait until professionalism, scores for the second optimal croppings output by the second DCNN 308 to exceed the professionalism scores for the first optimal croppings from the first DCNN 304. It may be enough to come "close enough" such that the benefits in saving computing resources exceed the decline in accuracy from using the second DCNN 304. As such, a formula may ⁇ be used by the DCNN selector 310 such as changing over when the following formula is deemed true:
  • This formula may be tweaked based on implementation. For example, if multiple (e.g. n) images are used in this testing process, the average differential between the first DCNN 304 path and second DCNN 308 path may ⁇ be used as follows: Professionalism Score Pirst : — Professionalism Score Second t + Distance Factor n
  • the DCNN selector 310 determines that it is time to changeover to the second DCNN 308.
  • one or more other metrics may be used.
  • this metric is a measure of how similar two separate outputs for different areas of the same input image are for the second DCNN 308. The more similar the outputs, the
  • one of these versions may be the left 75% of the image (with the right
  • the second of the versions may be the right 75% of the
  • DCNN 308 is "ready,” it should output the same optimal cropping for both of
  • FIG. 18 is a flow diagram illustrating a method 1800 of training and utilizing DCNNs, in accordance with an example embodiment. At operation
  • a first DCNN is trained to output a value for a first metric by inputting a
  • documents may have been labeled with a value for the first metric.
  • the documents described herein may be images, although this disclosure is not limited to embodiments where the documents are images.
  • a plurality of possible transformations of a first input document are fed to the first DCNN, obtaining a value for the first metric for each of the plurality of possible transformations.
  • a first transformation is selected from the plurality of possible transformations based on the values for the first
  • a second DCNN is trained to output a transformation for a document by inputting the selected first transformation and its corresponding value for the first metric to the second DCNN.
  • a second input document is fed to the second DCNN, obtaining a second transformation of the second input document.
  • FIG. 19 is a flow diagram illustrating a method 1900 for changing a system over from a first DCNN to a second DCNN, in accordance with an example embodiment.
  • a loop is begun for each of one or more input documents.
  • transformation of the input document is determined by passing the first transformation to the first DCNN.
  • a second transformation of the input document is obtained by passing the input document to the second DCNN.
  • the second transformation of the input document is passed to the first DCNN, obtaining a second value for the first metric for the second transformation of the input document.
  • the first and second transformations are of the same first transformation type.
  • a difference is determined between the first value and the second value.
  • the method 1900 loops back to operation 1902. If so, then at operation 1912 it is determined whether to change the system over from the first DCNN to the second DCNN, based on the difference between the first value and the second value for each of the one or more input documents. If it is determined that the changeover should not occur, the method loops back to operation 1902 for a different set of input documents. If it is determined that a changeover should occur, then at operation 1914, for transformations of the first transformation type, new input documents to the system are caused to be passed to the second DCNN in lieu of being passed to the first DCNN.
  • Example 1 A computerized method of training and utilizing deep convoiutional neural networks (DCNNs), the method comprising:
  • Example 2 The method of Example 1, wherein the sample documents and the first and second input documents are images.
  • Example 3 The method of Examples 1 or 2, wherein the inputting the plurality of sample documents to the first DCNN includes, for each sample document:
  • the convoiutional layer comprising one or more filters having dynamically adjustable weights, the one or more filters configured to filter the image to produce an output volume for the corresponding sample document, the output volume compri sing a different feature map for each of the one or more filters; passing the output volume from the convoiutional layer through a
  • nonlinearity layer the nonlinearity layer applying a nonlinearity function to the output volume from the convoiutional layer; passing the output volume from the nonlinearity layer through a pooling layer, the pooling layer lowering spatial dimensions of the output volume from the nonlinearity layer;
  • the classification layer comprising a specialized convolutional layer having a filter designed to output a professionalism score for the sample document based on the output volume from the pooling layer;
  • the loss layer applying a loss function to the sample document, resuitmg in an indication of a level of error in the professionalism score for the sample document from the classification layer in comparison to the professionalism score from the label of the sample document;
  • Example 4 The method of Example 3, wherein the feeding a plurality of possible transformations of the first input document to the first DCNN includes, for each of the plurality of possible transformations:
  • Example 5 The method of any of Examples 1-4, wherein the plurality of possible transformations are different croppings of an image.
  • Example 6 The method of any of Examples 1-5, further comprising:
  • Example 7 The method of any of Examples 1 -6, wherein the second DCNN comprises multiple stages, each stage containing a different convolutional layer, nonlinearity layer, and pooling layer.
  • Example 8 A system comprising:
  • a computer readable medium having instructions stored there on, which, when executed by a processor, cause the system to:
  • Example 9 The system of Example 8, wherein the sample documents and the first and second input documents are images.
  • Example 10 The system of Examples 8 or 9, wherein the inputting the plurality of sample documents to the first DCNN includes, for each sample document:
  • the convolutional layer comprising one or more filters having dynamically adjustable weights, the one or more filters configured to filter the image to produce an output volume for the corresponding sample document, the output volume comprising a diffe ent feature map for each of the one or more filters; passing the output volume from the convolutional layer through a nonlinearity layer, the nonlinearity layer applying a nonlinearity function to the output volume from the convolutional layer;
  • the classification layer comprising a specialized convolutional layer having a filter designed to output a professionalism score for the sample document based on the output volume from the pooling layer;
  • the loss layer applying a loss function to the sample document, resulting in an indication of a level of error in the professionalism score for the sample document from the classification layer in comparison to the professionalism score from the label of tlie sample document; determining whether a combination of the levels of error for the plurality of sample documents transgresses a preset threshold;
  • Example 11 The system of Example 10, wherein the feeding a plurality of possible transformations of the first input document to the first DCNN includes, for each of the plurality of possible transformations:
  • Example 12 The sy stem of any of Examples 8-11, wherein the plurality of possible transformations are different croppings of an image.
  • Example 13 The system of any of Examples 8-12, wherein the instructions further cause the system to:
  • Example 14 The system of any of Examples 8-13, wherein the second DCNN comprises multiple stages, each stage containing a different convolutional layer, nonlinearity layer, and pooling layer.
  • Example 15 A non-transitory machine-readable storage medium, comprising instructions, which when implemented by one or more machines, cause the one or more machines to perform operations comprising:
  • Example 16 The non-transitory machine-readable storage medium of Example 15, wherein the sample documents and the first and second input documents are images.
  • Example 17 The non-transitory machine-readable storage medium of
  • Examples 15 or 16, wherein the inputting the plurality of sample documents to the first DCNN includes, for each sample document:
  • the convolutional layer comprising one or more filters having dynamically adjustable weights, the one or more filters configured to filter the image to produce an output volume for the corresponding sample document, the output volume comprising a different feature map for each of the one or more filters; passing the output volume from the convolutional lay er through a nonlinearity layer, the nonlinearity layer applying a nonlinearity function to the output volume from the convolutional layer;
  • the classification layer comprising a specialized convolutional layer having a filter designed to output a professionalism score for the sample document based on the output volume from the pooling layer;
  • the loss layer applying a loss function to the sample document, resulting in an indication of a level of error in the professionalism score for the sample document from the classification layer in comparison to the professionalism score from the label of the sample document;
  • Example 18 The non -transitory machine-readable storage medium of Example
  • feeding a plurality of possible transformations of the first input document to the first DCNN includes, for each of the plurality of possible transformations:
  • Example 19 The non -transitory machine-readable storage medium of any of Examples 15-18, wherein the plurality of possible transformations are different croppings of an image.
  • Example 20 The non-transitory machine-readable storage medium of any of Examples 15-19, wherein the instructions further cause the one or more machines to perform operations comprising:
  • Example 21 A computerized method for changing a system over from a first deep convolutional neural network (DCNN) to a second DCNN, the first DCNN trained to output a value for a first metric for a document input to the first DCNN and the second DCNN trained to output a transformation of a first transformation type for a document input to the first DCNN, the method compr sing:
  • Example 22 The method of Example 21, wherein the determining whether to change the system over includes adding a distance factor to the difference, the distance factor set based on savings of computational power by virtue of using the second DCNN in lieu of the first DCNN.
  • Example 23 The method of Examples 21 or 22, wherein the first metric is a professionalism score.
  • Example 24 The method of any of Examples 21-23, wherein the determining whether to change the system over includes passing two versions of the same input document to the second DCNN and determining whether the second DCNN outputs the same second transformation for each of the two versions.
  • Example 25 Tire method of any of Examples 21-24, wherein the two versions of the same input document are partially overlapping portions of a single image.
  • Example 26 The method of any of Examples 21 -25, wherein the first transformation is obtained by
  • Example 27 The method of any of Examples 21-26, wherein the
  • transformation type is a cropping
  • Example 28 A system comprising:
  • a computer readable medium having instructions stored there on, which, when executed by a processor, cause the system to perform a method for changing a system over from a first deep convolutionai neural network (DCNN) to a second DCNN , the first DCNN trained to output a value for a first metric for a document input to the fi rst DCNN and the second DCNN trained to output a transformation of a first transformation type for a document input to the first DCNN, by:
  • DCNN deep convolutionai neural network
  • Example 29 The system of Example 28,wherein the determining whether to change the system, over includes adding a distance factor to the difference, the distance factor set based on savings of computational power by virtue of using the second DCNN in lieu of the first DCNN .
  • Example 30 The system of Examples 28 or 29, wherein the first metric is a professionalism score.
  • Example 3 The system of any of Examples 28-30, wherein the determining whether to change the system over includes passing two versions of the same input document to the second DCNN and determining whether the second DCNN outputs the same second transformation for each of the two versions.
  • Example 32 The system of any of Examples 28-31, wherein the two versions of the same input document are partially overlapping portions of a single image.
  • Example 33 The system of any of Examples 28-32, wherein the first
  • transfonnation is obtained by feeding a plurality of possible transformations of the first input document to the first DCNN and, for each of the plurality of possible transformations, within the first DCNN:
  • Example 34 The system of any of Examples 28-33, wherein the
  • transformation type is a cropping
  • Example 35 A non-transitory machine-readable storage medium comprising instructions, which when implemented by one or more machines, cause the one or more machines to perform operations for changing a system over from, a first deep convoiutional neural network (DCNN) to a second DCNN, the first DCNN trained to output a value for a first metric for a document input to the first DCNN and the second DCNN trained to output a transformation of a first transformation type for a document input to the first DCNN, the operations comprising:
  • DCNN deep convoiutional neural network
  • Example 36 The non-transitory machine-readable storage medium of Example 35,wherein the determining whether to change the system over includes adding a distance factor to the difference, the distance factor set based on savings of computational power by virtue of using the second DCNN in lieu of the first DCNN.
  • Example 37 The non-transitory machine-readable storage medium of
  • Example 38 The non-transitory machine -readable storage medium of any of Examples 35-37, wherein the determining whether to change the system over includes passing two versions of the same input document to the second DCNN and determining whether the second DCN N outputs the same second transformation for each of the two versions.
  • Example 39 The non-transitory machine-readable storage medium of any of Examples 35-38, wherein the two versions of the same input document are partially overlapping portions of a single image.
  • Example 40 The non-transitory machine-readable storage medium of any of Examples 35-39, wherein the first transformation is obtained by
  • Modules may constitute either software modules (e.g., code embodied on a machine-readable medium) or hardware modules.
  • a "hardware module " ' is a tangible unit capable of performing certain operations and may be configured or arranged in a certain physical manner.
  • one or more computer systems e.g., a standalone computer system, a client computer system, or a server computer system
  • one or more hardware modules of a computer system e.g., a processor or a group of processors
  • software e.g., an application or application portion
  • a hardware module may be implemented mechanically, electronically, or any suitable combination thereof.
  • a hardware module may include dedicated circuitry or logic that is permanently configured to perform, certain operations.
  • a hardware module may be a special-purpose processor, such as a Field-Programmable Gate Array (FPGA) or an Application Specific Integrated Circuit (ASIC).
  • a hardware module may also include programmable logic or circuitry that is temporarily configured by software to perform certain operations.
  • a hardware module may include software executed by a general-purpose processor or other programmable processor. Once configured by such software, hardware modules become specific machines (or specific components of a machine) uniquely tailored to perform the configured functions and are no longer general-purpose processors. It will be appreciated that the decision to implement a hardware module mechanically, in dedicated and permanently configured circuitry, or in temporarily configured circuitry (e.g., configured by software) may be driven by cost and time considerations.
  • hardware module ' should be understood to encompass a tangible entity, be that an entity that is physically constructed, permanently configured (e.g., hardwired), or temporarily configured (e.g., programmed) to operate in a certain manner or to perform certain operations described herein.
  • “hardware -implemented module” refers to a hardware module. Considering embodiments in which hardware modules are temporarily configured (e.g., programmed), each of the hardware modules need not be configured or instantiated at any one instance in time.
  • a hardware module comprises a general-purpose processor configured by software to become a special-purpose processor
  • the general-purpose processor may be configured as respectively different special-purpose processors (e.g., comprising different hardware modules) at different times.
  • Software accordingly configures a particular processor or processors, for example, to constitute a particular hardware module at one instance of tim e and to constitute a different hardware module at a different instance of time.
  • Hardware modules can provide information to, and receive information from, other hardware modules. Accordingly, the described hardware modules may be regarded as being communicatively coupled. Where multiple hardware modules exist contemporaneously, communications may be achieved through signal transmission (e.g., over appropriate circuits and buses) between or among two or more of the hardware modules. In embodiments in which multiple hardware modules are configured or instantiated at different times, communications between such hardware modules may be achieved, for example, through the storage and retrieval of information in memor - structures to which the multiple hardware modules have access. For example, one hardware module may perform an operation and store the output of that operation in a memory device to which it is communicatively coupled. A further hardware module may then, at a later time, access the memory device to retrieve and process the stored output. Hardware modules may also initiate
  • communications with input or output devices can operate on a resource (e.g., a collection of information).
  • a resource e.g., a collection of information
  • processors may be temporarily configured (e.g., by software) or permanently configured to perform the relevant operations. Whether temporarily or permanently configured, such processors may constitute processor-implemented modules that operate to perform one or more operations or functions described herein.
  • processor-implemented module refers to a hardware module implemented using one or more processors.
  • the methods described herein may be at least partially processor-implemented, with a particular processor or processors being an example of hardware.
  • a particular processor or processors being an example of hardware.
  • the operations of a method may be performed by one or more processors or processor-implemented modules.
  • the one or more processors may also operate to support performance of the relevant operations in a "cloud computing" environment or as a ' " software as a service” (SaaS).
  • SaaS software as a service
  • at least some of the operations may be performed by a group of computers (as examples of machines including processors), with these operations being accessible via a network (e.g., the Internet) and via one or more appropriate interfaces (e.g., an API).
  • the performance of certain of the operations may be distributed among the processors, not only residing within a single machine, but deployed across a number of machines.
  • the processors or processor-implemented modules may be located in a single geographic location (e.g., within a home environment, an office environment, or a server farm). In other example embodiments, the processors or processor-implemented modules may be distributed across a number of geographic locations.
  • FIGs. 1-19 The modules, methods, applications, and so forth described in conjunction with FIGs. 1-19 are implemented in some embodiments in the context of a machine and an associated software architecture.
  • the sections below describe representative software architecture(s) and machine (e.g., hardware) architecture(s) that are suitable for use with the disclosed
  • Software architectures are used in conjunction with hardware architectures to create devices and machines tailored to particular purposes. For example, a particular hardware architecture coupled with a particular software architecture will create a mobile device, such as a mobile phone, tablet device, or so forth. A slightly different hardware and software architecture may yield a smart device for use in the "internet of things," while yet another combination produces a server computer for use within a cloud computing architecture. Not ail combinations of such software and hardware architectures are presented here, as those of skill in the art can readily understand how to implement the inventive subject matter in different contexts from the disclosure contained herein.
  • FIG. 20 is a block diagram 2000 illustrating a representative software architecture 2002, which may be used in conjunction with various hardware architectures herein described.
  • FIG. 20 is merely a non-limiting example of a software architecture, and it will be appreciated that many other architectures may be implemented to facilitate the functionality described herein.
  • the software architecture 2002 may be executing on hardware such as a machine 1800 of FIG. 18 that includes, among other things, processors 1810,
  • a representative hardware layer 2004 is illustrated and can represent, for example, the machine 1800 of FIG. 18.
  • the representative hardware layer 2004 comprises one or more processing units 2006 having associated executable instructions 2008.
  • the executable instructions 2008 represent the executable instructions of the software architecture 2002, including implementation of the methods, modules, and so forth of FIGs. 1-19.
  • the hardware layer 2004 also includes memory and or storage modules 2010, which also have the executable instructions 2008.
  • the hardware layer 2004 may also comprise other hardware 2012, which represents any other hardware of the hardware layer 2004, such as the other hardware illustrated as part of the machine 1800.
  • the software architecture 2002 may be conceptualized as a stack of layers where each layer provides particular functionality.
  • the software architecture 2002 may include layers such as an operating system 2014, libraries 2016, frameworks/middleware 2018, applications 2020, and a presentation layer 2044, Operationally, the applications
  • the 2020 and/or other components within the layers may invoke API calls 2024 through the software stack and receive responses, returned values, and so forth, illustrated as messages 2026, in response to the API calls 2024.
  • the layers illustrated are representative in nature and not all software architectures have all layers. For example, some mobile or special purpose operating systems may not provide a layer of frameworks/middleware 2018, while others may provide such a layer. Other software architectures may include additional or different layers.
  • the operating system 2014 may manage hardware resources and provide common services.
  • the operating system 2014 may include, for example, a kernel 2028, services 2030, and drivers 2032.
  • the kernel 2028 may act as an abstraction layer between the hardware and the other software layers.
  • the kernel 2028 may be responsible for memory management, processor management (e.g., scheduling), component management, networking, security settings, and so on.
  • the services 2030 may provide other common services for the other software layers.
  • the drivers 2032 may be responsible for controlling or interfacing with the underlying hardware.
  • the drivers 2032 may include display drivers, camera drivers, Bluetooth® drivers, flash memory drivers, serial communication drivers (e.g., Universal Serial Bus (USB) drivers), Wi-Fi® drivers, audio drivers, power management drivers, and so forth depending on the hardware configuration.
  • USB Universal Serial Bus
  • the libraries 2016 may provide a common infrastructure that may be utilized by the applications 2020 and/or other components and/or layers.
  • the libraries 2016 typically provide functionality that allows other software modules to perform tasks in an easier fashion than by interfacing directly with the underlying operating system 2014 functionality (e.g., kernel 2028, services 2030, and/or drivers 2032).
  • the libraries 2016 may include system 2034 libraries (e.g., C standard library) that may provide functions such as memory allocation functions, string manipulation functions, mathematic functions, and the like.
  • libraries 2016 may include API 2036 libraries such as media libraries (e.g., libraries to support presentation and manipulation of various media formats such as MPEG4, H.264, MP3, AAC, AMR, JPG, PNG), graphics libraries (e.g., an OpenGL framework that may be used to render 2D and 3D graphic content on a display), database libraries (e.g., SQLite that may provide various relational database functions), web libraries (e.g., WebKit that may- provide web browsing functionality), and the like.
  • the libraries 2016 may also include a wide variety of other libraries 2038 to provide many other APIs to the applications 2020 and other software components/modules.
  • the frameworks 2018 may provide a higher-level common infrastructure that may be utilized by the applications 2020 and/or other software components/modules.
  • the frameworks 2018 may provide various graphic user interface (GUI) functions, high-level resource management, high-level location sen-ices, and so forth.
  • GUI graphic user interface
  • the frameworks 2018 may provide a broad spectrum of other APIs that may be utilized by the applications 2020 and/or other software components/modules, some of which may be specific to a particular operating system or platform.
  • the applications 2020 include buiit-m applications 2040 and/or third part ⁇ ' applications 2042.
  • Examples of representative built-in applications 2040 may include, but are not limited to, a contacts application, a browser application, a book reader application, a location application, a media application, a messaging application, and/or a game application.
  • the third party applications 2042 may include any of the built-in applications as well as a broad assortment of other applications.
  • the third party application 2042 e.g., an application developed using the AndroidTM or iOSTM software development kit (SDK) by an entity other than the vendor of the particular platform
  • the third party application 2042 may be mobile software running on a mobile operating system such as iOSTM, AndroidTM, Windows® Phone, or other mobile operating systems.
  • the third party application 2042 may invoke the API calls 2024 provided by the mobile operating system such as the operating system 2014 to facilitate functionality described herein .
  • the applications 2020 may utilize built-in operating system 2014 functions (e.g., kernel 2028, services 2030, and/or drivers 2032), libraries 2016
  • frameworks/middleware 2018 to create user interfaces to interact with users of the system.
  • interactions with a user may occur through a presentation layer, such as the presentation layer 2044.
  • the application/module "logic" can be separated from the aspects of the application/module that interact with a user.
  • a virtual machine creates a software environment where applications/modules can execute as if they were executing on a hardware machine (such as the machine 1800 of FIG. 18, for example).
  • a virtual machine is hosted by a host operating system (e.g., operating system 2014 in FIG. 20) and typically, although not always, has a virtual machine monitor 2046, which manages the operation of the virtual machine as well as the interface with the host operating system (e.g., operating system 2014).
  • a software architecture executes within the virtual machine 2048, such as an operating system 2,050, libraries 2052, frameworks/middleware 2054, applications 2056, and/or a presentation layer 2058. These layers of software architecture executing within the virtual machine 2048 can be the same as corresponding layers previously described or may be different.
  • FIG. 21 is a block diagram illustrating components of a machine 2100, according to some example embodiments, able to read instructions from a machine-readable medium (e.g., a machine-readable storage medium) and perform, any one or more of the methodologies discussed herein.
  • FIG. 21 shows a diagrammatic representation of the machine 2100 in the example form of a computer system, within which instructions 21 16 (e.g., software, a program, an application, an applet, an app, or other executable code) for causing the machine 2, 100 to perform any one or more of the methodologies discussed herein may be executed.
  • the instructions 21 16 transform the general, non- programmed machine into a particular machine programmed to cany out the described and illustrated functions in the manner described.
  • the machine 2100 operates as a standalone device or may be coupled (e.g., networked) to other machines.
  • the machine 2100 may operate in the capacity of a server machine or a client machine in a server-client network environment, or as a peer machine in a peer- to-peer (or distributed) network environment.
  • the machine 2100 may comprise, but not be limited to, a server computer, a client computer, a PC, a tablet computer, a laptop computer, a netbook, a set-top box (STB), a personal digital assistant (PDA), an entertainment media system, a cellular telephone, a smart phone, a mobile device, a wearable device (e.g., a smart watch), a smart home device (e.g., a smart appliance), other smart devices, a web appliance, a network router, a network switch, a network bridge, or any machine capable of executing the instructions 2116, sequentially or otherwise, that specify actions to be taken by the machine 2100.
  • the term "machine” shall also be taken to include a collection of machines 2100 that individually or jointly execute the instructions 2116 to perform any one or more of the methodologies discussed herein.
  • the machine 2100 may include processors 2110, memory/storage 2130, and I/O components 2150, which may be configured to communicate with each other such as via a bus 2102.
  • the processors 21 10 e.g., a Central Processing Unit (CPU), a Reduced Instruction Set Computing (RISC) processor, a Complex Instruction Sec Computing (CISC) processor, a Graphics Processing Unit (GPU), a Digital Signal Processor (DSP), an ASIC, a Radio-Frequency Integrated Circuit (RFIC), another processor, or any suitable combination thereof
  • the processors 21 10 may include, for example, a processor 2112 and a processor 2114 that may execute the instructions 2116.
  • RISC Reduced Instruction Set Computing
  • CISC Complex Instruction Sec Computing
  • GPU Graphics Processing Unit
  • DSP Digital Signal Processor
  • RFIC Radio-Frequency Integrated Circuit
  • processor is intended to include multi-core processors that may comprise two or more independent processors (sometimes referred to as “cores”) that may execute instructions contemporaneously.
  • FIG. 21 shows multiple processors 2110, the machine 2100 may include a single processor with a single core, a single processor with multiple cores (e.g., a multi-core processor), multiple processors with a single core, multiple processors with multiples cores, or any combination thereof.
  • the memory/storage 2130 may include a memory 2132, such as a main memor ', or other memory storage, and a storage unit 2136, both accessible to the processors 21 10 such as via the bus 2102.
  • the storage unit 2136 and memory 2132 store the instructions 2116 embodying any one or more of the methodologies or functions described herein.
  • the instructions 21 16 may also reside, completely or partially, within the memory 2132, within the storage unit
  • machine-readable medium means a device able to store instructions and data temporarily or permanently and may include, but is not limited to, random-access memory (RAM), read-only memory (ROM), buffer memory, flash memor ', optical media, magnetic media, cache memory-, other types of storage (e.g., Erasable Programmable Read-Only Memory (EEPROM)), and/or any suitable combination thereof.
  • RAM random-access memory
  • ROM read-only memory
  • EEPROM Erasable Programmable Read-Only Memory
  • machine- readable medium should be taken to include a single medium or multiple media (e.g., a centralized or distributed database, or associated caches and sen/ers) able to store the instructions 2116.
  • the term “machine-readable medium” shall also be taken to include any medium, or combination of multiple media, that is capable of storing instructions (e.g., instructions 21 16) for execution by a machine (e.g., machine 2100), such that the instructions, when executed by one or more processors of the machine (e.g., processors 2110), cause the machine to perform any one or more of the methodologies described herein.
  • a “machine-readable medium” refers to a single storage apparatus or device, as well as “cloud-based” storage systems or storage networks that include multiple storage apparatus or devices.
  • the term “machine-readable medium” excludes signals per se.
  • the I/O components 2150 may include a wide variety of components to recei ve input, pro vide output, produce output, transmit information, exchange information, capture measurements, and so on.
  • I/O components 2150 that are included in a particular machine will depend on the type of machine. For example, portable machines such as mobile phones will likely include a touch input device or other such input mechanisms, while a headless server machine will likely not include such a touch input device. It will be appreciated that the I/O components 2150 may include many other components that are not shown in FIG. 21. The I/O components 2150 are grouped according to functionality merely for simplifying the following discussion and the grouping is in no way limiting. In various example embodiments, the I/O components
  • the 2150 may include output components 2152 and input components 2154.
  • the output components 2152 may include visual components (e.g., a display such as a plasma display panel (PDP), a light emitting diode (LED) display, a liquid crystal display (LCD), a projector, or a cathode ray tube (CRT)), acoustic components (e.g., speakers), haptic components (e.g., a vibratory motor, resistance mechanisms), other signal generators, and so forth.
  • a display such as a plasma display panel (PDP), a light emitting diode (LED) display, a liquid crystal display (LCD), a projector, or a cathode ray tube (CRT)
  • acoustic components e.g., speakers
  • haptic components e.g., a vibratory motor, resistance mechanisms
  • the input components 2154 may include alphanumeric input components (e.g., a keyboard, a touch screen configured to receive alphanumeric input, a photo- optical keyboard, or other alphanumeric input components), point based input components (e.g., a mouse, a touchpad, a trackball, a joy stick, a motion sensor, or another pointing instrument), tactile input components (e.g., a physical button, a touch screen that provides location and/or force of touches or touch gestures, or other tactile input components), audio input components (e.g., a microphone), and the like.
  • alphanumeric input components e.g., a keyboard, a touch screen configured to receive alphanumeric input, a photo- optical keyboard, or other alphanumeric input components
  • point based input components e.g., a mouse, a touchpad, a trackball, a joy stick, a motion sensor, or another pointing instrument
  • tactile input components e.g., a physical button, a
  • the I/O components 2150 may include biometric components 2156, motion components 2158, environmental components 2160, or position components 2162, among a wide array of other components.
  • the biometric components 2156 may include components to detect expressions (e.g., hand expressions, facial expressions, vocal expressions, body gestures, or eye tracking), measure biosignals (e.g., blood pressure, heart rate, body temperature, perspiration, or brain waves), identify a person (e.g., voice identification, retinal identification, facial identification, fingerprint identification, or electroencephalogram based identification), and the like.
  • the motion components 2158 may include acceleration sensor components (e.g., accelerometer), gravitation sensor components, rotation sensor components (e.g., gyroscope), and so forth.
  • the environmental components 2160 may include, for example, illumination sensor components (e.g., photometer), temperature sensor components (e.g., one or more thermometers that detect ambient temperature), humidity sensor components, pressure sensor components (e.g., barometer), acoustic sensor components (e.g., one or more microphones that detect background noise), proximity sensor components (e.g., infrared sensors that detect nearby objects), gas sensors (e.g., gas detection sensors to detect concentrations of hazardous gases for safety or to measure pollutants in the atmosphere), or other components that may provide indications, measurements, or signals
  • illumination sensor components e.g., photometer
  • temperature sensor components e.g., one or more thermometers that detect ambient temperature
  • humidity sensor components e.g., pressure sensor components (e.g., barometer)
  • acoustic sensor components
  • the 2162 may include location sensor components (e.g., a Global Position System (GPS) receiver component), altitude sensor components (e.g., altimeters or barometers that detect air pressure from which altitude may be deri ved), orientation sensor components (e.g., magnetometers), and the like.
  • location sensor components e.g., a Global Position System (GPS) receiver component
  • altitude sensor components e.g., altimeters or barometers that detect air pressure from which altitude may be deri ved
  • orientation sensor components e.g., magnetometers
  • the I/O components 2150 may include communication components 2164 operable to couple the machine 2100 to a network 2180 or devices 2170 via a coupling 2, 182 and a coupling 2172, respectively.
  • the communication components 2164 may include a network interface component or other suitable device to interface with the network 2180.
  • the communication components 2164 may include wired communication components, wireless communication components, cellular communication components, Near Field Communication (NFC) components, Bluetooth® components (e.g., Bluetooth® Low Energy), Wi-Fi® ) components, and other communication components to provide communication via other modalities.
  • the devices 2170 may be another machine or any of a wide variety of peripheral devices (e.g., a peripheral device coupled via a USB).
  • the communication components 2164 may detect identifiers or include components operable to detect identifiers.
  • the communication components 2164 may include Radio Frequency
  • RFID Identification
  • NFC smart tag detection components optical reader components (e.g., an optical sensor to detect one- dimensional bar codes such as Universal Product Code (UPC) bar code, multidimensional bar codes such as Quick Response (QR) code, Aztec code. Data Matrix, Dataglyph, MaxiCode, PDF417, Ultra Code, UCC RSS-2D ba code, and other optical codes), or acoustic detection components (e.g., microphones to identify tagged audio signals).
  • UPC Universal Product Code
  • QR Quick Response
  • acoustic detection components e.g., microphones to identify tagged audio signals.
  • a variety of information may be derived via the communication components 2164, such as location via Internet Protocol (IP) geoiocation, location via Wi-Fi® signal triangulation, location via detecting an NFC beacon signal that may indicate a particular location, and so forth.
  • IP Internet Protocol
  • Wi-Fi® Wireless Fidelity
  • one or more portions of the network 2180 may be an ad hoc network, an intranet, an extranet, a virtual private network (VPN), a local area network (LAN), a wireless LAN (WLAN), a WAN, a wireless WAN (WW AN), a metropolitan area network (MAN), the Internet, a portion of the Internet, a portion of the Public Switched Telephone Network (PSTN , a plain old telephone service (POTS) network, a cellular telephone network, a wireless network, a Wi-Fi® network, another type of network, or a combination of two or more such networks.
  • VPN virtual private network
  • LAN local area network
  • WLAN wireless LAN
  • WAN wide area network
  • WW AN wireless WAN
  • MAN metropolitan area network
  • PSTN Public Switched Telephone Network
  • POTS plain old telephone service
  • the network 2180 or a portion of the network 2180 may include a wireless or cellular network and the coupling 2182 may be a Code Division Multiple Access (CDMA) connection, a Global System for Mobile communications (GSM) connection, or another type of cellular or wireless coupling.
  • the coupling 2182 may implement any of a variety of types of data transfer technology, such as Single Carrier Radio Transmission Technology (lxRTT), Evolution -Data Optimized (EVDO) technology.
  • GPRS General Packet Radio Sendee
  • EDGE Enhanced Data rates for GSM Evolution
  • 3GPP Third Generation Partnership Project
  • 4G fourth generation wireless (4G) networks
  • High Speed Packet Access (HSPA) Worldwide
  • WiMAX Microwave Access
  • LIE Long Term Evolution
  • the instructions 2116 may be transmitted or received over the network 2180 using a transmission medium via a network interface device (e.g., a network interface component included in the communication components 2164) and utilizing any one of a number of well-known transfer protocols (e.g., HTTP). Similarly, the instructions 2116 may be transmitted or received using a transmission medium via the coupling 2 72 (e.g., a peer-to-peer coupling) to the devices 2170.
  • the term "transmission medium” shall be taken to include any intangible medium that is capable of storing, encoding, or carrying the instructions 2116 for execution by the machine 2100, and includes digital or analog communications signals or other intangible media to facilitate communication of such software.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Image Analysis (AREA)

Abstract

Un mode de réalisation de l'invention donné à titre d'exemple concerne un premier DCNN qui est entraîné pour produire une valeur pour une première mesure en entrant une pluralité de documents d'échantillons dans le premier DCNN, chacun des documents d'échantillons ayant été étiqueté avec une valeur pour la première mesure. Ensuite, une pluralité de transformations possibles d'un premier document d'entrée sont fournies au premier DCNN, ce qui permet d'obtenir une valeur pour la première mesure pour chaque transformation possible de la pluralité de transformations possibles. Une première transformation est sélectionnée parmi la pluralité de transformations possibles en fonction des valeurs pour la première mesure pour chaque transformation possible de la pluralité de transformations possibles. Ensuite, un second DCNN est entraîné pour produire une transformation pour un document en entrant la première transformation sélectionnée dans le second DCNN. Le second document d'entrée est fourni au second DCNN, ce qui permet d'obtenir une seconde transformation du second document d'entrée.
PCT/US2017/034972 2016-05-31 2017-05-30 Entraînement d'un réseau neuronal à l'aide d'un autre réseau neuronal Ceased WO2017210174A1 (fr)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201780033295.4A CN109564636B (zh) 2016-05-31 2017-05-30 使用一个神经网络来训练另一个神经网络

Applications Claiming Priority (4)

Application Number Priority Date Filing Date Title
US15/169,063 US10380480B2 (en) 2016-05-31 2016-05-31 Changeover from one neural network to another neural network
US15/168,750 2016-05-31
US15/169,063 2016-05-31
US15/168,750 US11188823B2 (en) 2016-05-31 2016-05-31 Training a neural network using another neural network

Publications (1)

Publication Number Publication Date
WO2017210174A1 true WO2017210174A1 (fr) 2017-12-07

Family

ID=60478961

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/US2017/034972 Ceased WO2017210174A1 (fr) 2016-05-31 2017-05-30 Entraînement d'un réseau neuronal à l'aide d'un autre réseau neuronal

Country Status (2)

Country Link
CN (1) CN109564636B (fr)
WO (1) WO2017210174A1 (fr)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110705756A (zh) * 2019-09-07 2020-01-17 创新奇智(重庆)科技有限公司 一种基于输入凸神经网络的电力能耗优化控制方法

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110134852B (zh) * 2019-05-06 2021-05-28 北京四维图新科技股份有限公司 一种文档的去重方法、设备及可读介质
CN114730318A (zh) * 2019-11-28 2022-07-08 三菱电机株式会社 信息处理装置、信息处理方法以及信息处理程序
CN112541574B (zh) * 2020-12-03 2022-05-17 支付宝(杭州)信息技术有限公司 保护隐私的业务预测方法及装置
CN114003735B (zh) * 2021-12-24 2022-03-18 北京道达天际科技有限公司 基于情报文档面向知识图谱问答的实体消歧方法

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20090210218A1 (en) * 2008-02-07 2009-08-20 Nec Laboratories America, Inc. Deep Neural Networks and Methods for Using Same
US20160034788A1 (en) * 2014-07-30 2016-02-04 Adobe Systems Incorporated Learning image categorization using related attributes
US20160035078A1 (en) * 2014-07-30 2016-02-04 Adobe Systems Incorporated Image assessment using deep convolutional neural networks
US20160055409A1 (en) * 2014-08-19 2016-02-25 Qualcomm Incorporated Knowledge-graph biased classification for data
US20160140438A1 (en) * 2014-11-13 2016-05-19 Nec Laboratories America, Inc. Hyper-class Augmented and Regularized Deep Learning for Fine-grained Image Classification

Family Cites Families (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP0532053A2 (fr) * 1991-09-12 1993-03-17 Fuji Photo Film Co., Ltd. Procédé de reconnaissance d'images d'objets et procédé d'apprentissage pour un réseau neuronal
CN101795344B (zh) * 2010-03-02 2013-03-27 北京大学 数字全息图像压缩、解码方法及系统、传输方法及系统
CN101930561A (zh) * 2010-05-21 2010-12-29 电子科技大学 一种基于N-Gram分词模型的反向神经网络垃圾邮件过滤装置
US9406017B2 (en) * 2012-12-24 2016-08-02 Google Inc. System and method for addressing overfitting in a neural network
CN104112263B (zh) * 2014-06-28 2018-05-01 南京理工大学 基于深度神经网络的全色图像与多光谱图像融合的方法
CN105120130B (zh) * 2015-09-17 2018-06-29 京东方科技集团股份有限公司 一种图像升频系统、其训练方法及图像升频方法
CN105426861B (zh) * 2015-12-02 2019-05-21 百度在线网络技术（北京）有限公司 车道线确定方法及装置

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20090210218A1 (en) * 2008-02-07 2009-08-20 Nec Laboratories America, Inc. Deep Neural Networks and Methods for Using Same
US20160034788A1 (en) * 2014-07-30 2016-02-04 Adobe Systems Incorporated Learning image categorization using related attributes
US20160035078A1 (en) * 2014-07-30 2016-02-04 Adobe Systems Incorporated Image assessment using deep convolutional neural networks
US20160055409A1 (en) * 2014-08-19 2016-02-25 Qualcomm Incorporated Knowledge-graph biased classification for data
US20160140438A1 (en) * 2014-11-13 2016-05-19 Nec Laboratories America, Inc. Hyper-class Augmented and Regularized Deep Learning for Fine-grained Image Classification

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110705756A (zh) * 2019-09-07 2020-01-17 创新奇智(重庆)科技有限公司 一种基于输入凸神经网络的电力能耗优化控制方法
CN110705756B (zh) * 2019-09-07 2023-05-12 创新奇智(重庆)科技有限公司 一种基于输入凸神经网络的电力能耗优化控制方法

Also Published As

Publication number Publication date
CN109564636B (zh) 2023-05-02
CN109564636A (zh) 2019-04-02

Similar Documents

Publication Publication Date Title
US9904871B2 (en) Deep convolutional neural network prediction of image professionalism
US20170300811A1 (en) Dynamic loss function based on statistics in loss layer of deep convolutional neural network
US11188823B2 (en) Training a neural network using another neural network
US10990899B2 (en) Deep and wide machine learned model for job recommendation
US10043240B2 (en) Optimal cropping of digital image based on professionalism score of subject
US10380458B2 (en) Cold start machine learning algorithm
US10565562B2 (en) Hashing query and job posting features for improved machine learning model performance
US10043254B2 (en) Optimal image transformation based on professionalism score of subject
US11574005B2 (en) Client application content classification and discovery
CN111582297B (zh) 细粒度分类
US10726025B2 (en) Standardized entity representation learning for smart suggestions
US11372940B2 (en) Embedding user categories using graphs for enhancing searches based on similarities
US10984365B2 (en) Industry classification
US11204973B2 (en) Two-stage training with non-randomized and randomized data
US20170300862A1 (en) Machine learning algorithm for classifying companies into industries
US20190019157A1 (en) Generalizing mixed effect models for personalizing job search
US10380480B2 (en) Changeover from one neural network to another neural network
US10726355B2 (en) Parent company industry classifier
US20210319033A1 (en) Learning to rank with alpha divergence and entropy regularization
CN109564636B (zh) 使用一个神经网络来训练另一个神经网络
US20200380407A1 (en) Generalized nonlinear mixed effect models via gaussian processes
US20190332569A1 (en) Integrating deep learning into generalized additive mixed-effect (game) frameworks
US11194877B2 (en) Personalized model threshold
US10572835B2 (en) Machine-learning algorithm for talent peer determinations
US20190095531A1 (en) Job flavor faceting

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 17807323

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 17807323

Country of ref document: EP

Kind code of ref document: A1