A video translation system that generates an output video in a target language which includes a translated/output audio track that runs in synchrony with the video content of a received input video in a source language and further displays translated subtitles corresponding to the translated audio track is disclosed. Upon receiving the input video, the domain of the input video can be identified. A translation engine and a transcription engine are selected based on the domain and the pair of languages corresponding to the input video and the output video. The output audio track is generated using the translation engine and merged with a manipulated video, which runs in synchrony with the output audio track to generate the output video. The transcription engine generates subtitles translated from the source language to the target language for the output video.

In one aspect, the present disclosure relates to a method for reducing fraud in computer networks, the method including receiving, from each of a plurality of user devices, a request to block an ad displayed within a web browser installed on the user device, the request comprising image data and a forwarding URL associated with the ad; storing crowdsourced ad blocking data based on the received requests to block ads; receiving a request for a list of blocked ads; generating a list of blocked ads based on analyzing the crowdsourced ad blocking data, wherein analyzing the crowdsourced ad blocking data comprises identifying ads blocked by at least a threshold number of users; and sending the list of blocked ads to a first user device, the first user device comprising a browser extension configured to prevent ads within the list of blocked ads from being rendered in a browser.

A method includes executing an encoder machine learning model on multiple token values contained in a document to create an encoder hidden state vector. A decoder machine learning model executing on the encoder hidden state vector generates raw text comprising an entity value and an entity label for each of multiple entities. The method further includes generating a structural representation of the entities directly from the raw text and outputting the structural representation of the entities of the document.

A computing system generates a training data set for training the prediction model to detect defects present in a target surface of a target specimen and training the prediction model to detect defects present in the target surface of the target specimen based on the training data set. The computing system generates the training data set by identifying a set of images for training the prediction model, the set of images comprising a first subset of images. A deep learning network generates a second subset of images for subsequent labelling based on the set of images comprising the first subset of images. The deep learning network generates a third subset of images for labelling based on the set of images comprising the first subset of images and the labeled second subset of images. The computing system continues the process until a threshold number of labeled images is generated.

A deep learning based compression (DLBC) system applies trained models to compress binary code of an input image to a target codelength. For a set of binary codes representing the quantized coefficents of an input image, the DLBC system applies a first model that is trained to predict feature probabilities based on the context of each bit of the binary codes. The DLBC system compresses the binary code via adaptive arithmetic coding based on the determined probability of each bit. The compressed binary code represents a balance between a reconstruction quality of a reconstruction of the input image and a target compression ratio of the compressed binary code.

Examples of techniques for interactive generation of labeled data and training instances are provided. According to one or more embodiments of the present invention, a computer-implemented method for interactive generation of labeled data and training instances includes presenting, by the processing device, control labeling options to a user. The method further includes selecting, by a user, one or more of the presented control labeling options. The method further includes selecting, by a processing device, a representative set of unlabeled data samples based at least in part on the control labeling options selected by the user. The method further includes generating, by a processing device, a set of suggested labels for each of the unlabeled data samples.

Today, artificial neural networks are trained on large sets of manually tagged images. Generally, for better training, the training data should be as large as possible. Unfortunately, manually tagging images is time consuming and susceptible to error, making it difficult to produce the large sets of tagged data used to train artificial neural networks. To address this problem, the inventors have developed a smart tagging utility that uses a feature extraction unit and a fast-learning classifier to learn tags and tag images automatically, reducing the time to tag large sets of data. The feature extraction unit and fast-learning classifiers can be implemented as artificial neural networks that associate a label with features extracted from an image and tag similar features from the image or other images with the same label. Moreover, the smart tagging system can learn from user adjustment to its proposed tagging. This reduces tagging time and errors.

A computing system generates a training data set for training the prediction model to detect defects present in a target surface of a target specimen and training the prediction model to detect defects present in the target surface of the target specimen based on the training data set. The computing system generates the training data set by identifying a set of images for training the prediction model, the set of images comprising a first subset of images. A deep learning network generates a second subset of images for subsequent labelling based on the set of images comprising the first subset of images. The deep learning network generates a third subset of images for labelling based on the set of images comprising the first subset of images and the labeled second subset of images. The computing system continues the process until a threshold number of labeled images is generated.

An information extraction system divides learning data items into main clusters by performing clustering on a set of the learning data items for use in generation of clustering models that are information extraction models for extracting information from invoice data and generates the different information extraction models for the different main clusters by performing learning using the learning data items for the individual main clusters.

Examples of techniques for adaptive model training are provided. According to one or more embodiments of the present invention, a computer-implemented method for adaptive model training includes generating, by a processing system, a training instance based at least in part on a plurality of images that match a contextual specification of a target visual domain. The method further includes extracting, by the processing system, objects from one of the plurality of images. The method further includes for each extracted object, generating, by the processing system, a plurality of machine learning model features and label recommendations for a user.