Abstract: A machine learning model, such as a neural network, is partially encrypted with a homomorphic encryption scheme. Application of the machine learning model to data includes performing operations on plaintext and encrypting results of operations for input to other operations that are performed over cyphertext. Ciphertext output of such operations can be provided to a service that is able to decrypt the ciphertext output.

Abstract: Techniques for performing visual clustering with a hierarchical graph neural network framework including a joint linkage prediction and density estimation graph model are described. Embodiments herein recurrently run the joint linkage prediction and density estimation graph model to generate intermediate clusters in multiple iterations (e.g., until convergence) to obtain a final clustering result. In certain embodiments, for each iteration, the input graph contains nodes that are merged from nodes assigned to intermediate clusters from the previous iteration. By using a small and fixed bandwidth k in each iteration, embodiments herein alleviate the sensitivity to the k selection for different clustering applications. Certain embodiments herein remove the tuning of a different k (e.g., k-bandwidth) for k-nearest neighbor graph construction over different clustering applications.

Abstract: An image may be evaluated by a computer vision system to determine whether it is fit for analysis. The computer vision system may generate an embedding of the image. An embedding quality score (EQS) of the image may be determined based on the image's embedding and a reference embedding associated with a cluster of reference noisy images. The quality of the image may be evaluated based on the EQS of the image to determine whether the quality meets filter criteria. The image may be further processed when the quality is sufficient, or otherwise the image may be removed.

Abstract: Techniques for data synthesis for training datasets for machine learning applications are described. A first image of at least an object from a first viewpoint is obtained. The first image having associated first image metadata including a first location of a feature of the object in the first image. A model is generated from the first image, the model including a three-dimensional representation of the object. A second image is generated from the model, the second image including the object from a second viewpoint that is different from the first viewpoint. Second image metadata is generated, the second image metadata including a second location of the feature of the object in the second image, the second location corresponding to the first location adjusted for the difference between the second viewpoint and the first viewpoint.

Abstract: Techniques for building a backward compatible and backfill-free image search system are described. According to some embodiments, a backwards compatible training system trains a new embedding model to be backward compatible with the face embeddings (e.g., floating-point vectors) generated by a previous embedding model. In one embodiment, backwards compatible training uses a classifier of the previous embedding model as a form of constraint in the training of the new embedding model.

Abstract: Time domain action detecting methods and systems, electronic devices, and computer storage medium are provided. The method includes: obtaining a time domain interval in a video with an action instance and at least one adjacent segment in the time domain interval; separately extracting action features of at least two video segments in candidate segments, where the candidate segments comprises video segment corresponding to the time domain interval and adjacent segments thereof; pooling the action features of the at least two video segments in the candidate segments, to obtain a global feature of the video segment corresponding to the time domain interval; and determining, based on the global feature, an action integrity score of the video segment corresponding to the time domain interval. The embodiments of the present disclosure benefit accurately determining whether a time domain interval comprises an integral action instance, and improve the accuracy rate of action integrity identification.

Abstract: Time domain action detecting methods and systems, electronic devices, and computer storage medium are provided. The method includes: obtaining a time domain interval in a video with an action instance and at least one adjacent segment in the time domain interval; separately extracting action features of at least two video segments in candidate segments, where the candidate segments comprises video segment corresponding to the time domain interval and adjacent segments thereof; pooling the action features of the at least two video segments in the candidate segments, to obtain a global feature of the video segment corresponding to the time domain interval; and determining, based on the global feature, an action integrity score of the video segment corresponding to the time domain interval. The embodiments of the present disclosure benefit accurately determining whether a time domain interval comprises an integral action instance, and improve the accuracy rate of action integrity identification.