A method for generating labeling data is disclosed that describes an image content of images depicting at least one scene, wherein in a processing device image data are received from an imaging and a segmentation unit that detects at least one object in the image data. A graphical processing unit generates a respective graphical object marker that marks the at least one detected object and a display control unit displays an overlay of the at least one scene and the at least one object marker. An input reception unit receives a respective user input for each object marker, wherein the respective user input provides the image content of the image region marked by the object marker.

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for learning template representation libraries. In one aspect, a method includes obtaining an image depicting a physical environment, where the environment includes a given physical object. When possible, a position of the given object in the environment is inferred based on a template representation library using template matching techniques. In response to determining that the position of the given object in the environment cannot be inferred based on the template representation library using template matching techniques, the template representation library is automatically augmented with new template representations.

A multi-dimensional task facial beauty prediction method and system, and a storage medium are disclosed. The method includes the steps of: at a training phase, using first facial images to optimize a shared feature extraction network for extracting shared features and to train a plurality of sub-task networks for performing facial beauty classification tasks; at a testing phase, extracting shared features of second facial images; inputting the shared features to the trained plurality of sub-task networks; and obtaining a first beauty prediction result based on first output results of the plurality of sub-task networks.

A learning assistance device acquires a plurality of learned discriminators obtained by causing learning discriminators provided in a plurality of respective terminal devices to perform learning using image correct answer data, acquires a plurality of discrimination results obtained by causing a plurality of learned discriminators to discriminate the same input image, determines the correct answer data of the input image on the basis of the plurality of discrimination results, causes the discriminator to perform learning the input image and the correct answer data, and outputs a result thereof as a new learning discriminator to each terminal device.

A method includes accessing a structured content item from a first database and event data from a second database, the event data including sets of event attributes in a multi-dimensional namespace and associated with a respective point in time; determining a relevancy profile characterizing a metric of relevancy of the structured content item over a respective time interval, the metric of relevancy including a distance in the multi-dimensional namespace between attributes associated with the structured content and the sets of event attributes; generating, using the relevancy profile, second digital content including a subset of the structured content item; and providing the second digital content for rendering on a device. Related apparatus, systems, techniques and articles are also described.

A computer-implemented method of identifying filters for use in determining explainability of a trained neural network. The method comprises obtaining a dataset comprising the input image and an annotation of an input image, the annotation indicating at least one part of the input image which is relevant for inferring classification of the input image, determining an explanation filter set by iteratively: selecting a filter of the plurality of filters; adding the filter to the explanation filter set; computing an explanation heatmap for the input image by resizing and combining an output of each filter in the explanation filter set to obtain the explanation heatmap, the explanation heatmap having a spatial resolution of the input image; and computing a similarity metric by comparing the explanation heatmap to the annotation of the input image; until the similarity metric is greater than or equal to a similarity threshold; and outputting the explanation filter set.

Aspects of the disclosure relate to anomaly detection in cybersecurity training modules. A computing platform may receive information defining a training module. The computing platform may capture a plurality of screenshots corresponding to different permutations of the training module. The computing platform may input, into an auto-encoder, the plurality of screenshots corresponding to the different permutations of the training module, wherein inputting the plurality of screenshots corresponding to the different permutations of the training module causes the auto-encoder to output a reconstruction error value. The computing platform may execute an outlier detection algorithm on the reconstruction error value, which may cause the computing platform to identify an outlier permutation of the training module. The computing platform may generate a user interface comprising information identifying the outlier permutation of the training module. The computing platform may send the user interface to at least one user device.

This document relates to training of machine learning models. One example method involves providing a machine learning model having one or more mapping layers. The one or more mapping layers can include at least a first mapping layer configured to map components of pretraining examples into first representations in a space. The example method also includes performing a pretraining stage on the one or more mapping layers using the pretraining examples. The pretraining stage can include adding noise to the first representations of the components of the pretraining examples to obtain noise-adjusted first representations. The pretraining stage can also include performing a self-supervised learning process to pretrain the one or more mapping layers using at least the first representations of the training data items and the noise-adjusted first representations of the training data items.

A method for recognizing a key time point in a video includes: obtaining at least one video segment by processing each image frame in the video by an image classification model; determining a target video segment in the at least one video segment based on a shot type; obtaining respective locations of a first object and a second object in an image frame of the target video segment by an image detection model; and based on a distance between the location of the first object and the location of the second object in the image frame satisfying a preset condition, determining a time point of the image frame as the key time point of the video.

Aspects of the disclosure relate to monitoring devices at enterprise locations using machine-learning models to protect enterprise-managed information and resources. In some embodiments, a computing platform may receive, from one or more data source computer systems, passive monitoring data. Based on applying a machine-learning classification model to the passive monitoring data received from the one or more data source computer systems, the computing platform may determine to trigger a data capture process at an enterprise center. In response to determining to trigger the data capture process, the computing platform may initiate an active monitoring process to capture event data at the enterprise center. Thereafter, the computing platform may generate one or more alert messages based on the event data captured at the enterprise center. Then, the computing platform may send the one or more alert messages to one or more enterprise computer systems.