Google Inc.

Yair Movshovitz-Attias

Elad Edwin Tzvi Eban

Sergey Ioffe

Alexander Toshev

Saurabh Singh

King Hong Leung

Systems and methods described herein may relate to training an artificial neural network (ANN) using a differentiable Jaccard Loss approximation. An example embodiment may involve obtaining a training image and a corresponding ground truth mask that represents a desired segmentation of the training image. The embodiment may further involve applying an ANN on the training image to generate an output segmentation of the training image that depends on a plurality of weights of the ANN and determining a differentiable Jaccard Loss approximation based on the output segmentation of the training image and the ground truth mask. The embodiment also involves modifying one or more weights of the ANN based on the differentiable Jaccard Loss approximation and providing a representation of the ANN as modified to a mobile computing device.

Differentiable jaccard loss approximation for training an artificial neural network

The present disclosure provides systems and methods that enable distance metric learning using proxies. A machine-learned distance model can be trained in a proxy space in which a loss function compares an embedding provided for an anchor data point of a training dataset to a positive proxy and one or more negative proxies, where each of the positive proxy and the one or more negative proxies serve as a proxy for two or more data points included in the training dataset. Thus, each proxy can approximate a number of data points, enabling faster convergence. According to another aspect, the proxies of the proxy space can themselves be learned parameters, such that the proxies and the model are trained jointly. Thus, the present disclosure enables faster convergence (e.g., reduced training time). The present disclosure provides example experiments which demonstrate a new state of the art on several popular training datasets.

Distance metric learning using proxies

Audio recorded by a cellphone or other portable recording device (e.g., audio recorded as part of a video recording of a play or other event) is often of low quality, due to the limitations of the portable recording device. Multiple audio recordings, made during the same period of time and near the same location, can be combined to generate an improved-quality audio recording of an event. The audio recordings may be accessible to a server that selects the audio recordings and performs the combination. To protect the privacy of persons whose audio is used, more than a minimum number of recordings could be combined and/or no more than a threshold amount of any one recording could be used to generate a combined recording. Additionally, a provided 'clean' recording could include more than a threshold amount of the audio provided by a user or device that requests such a 'clean' recording.

Growing community of inventors

Mountain View, CA, United States of America

Yair Movshovitz-Attias