GTC 2020 S22192
Presenters: Shalini De Mello,NVIDIA
Abstract
Deep networks designed to observe human behavior that are trained on one set of subjects often do not perform optimally on others. Gaze-estimation is one such problem, where anatomical variations between subjects limit the performance of cross-person networks. Personalizing neural networks for each person on devices on the edge is the key to obtaining the best performance in such scenarios. However, this entails collecting thousands of training samples per person at the deployment site, which is simply not viable. To solve this, we present a novel and effective algorithm for training gaze networks with very few (less than 10) training examples per subject to create highly accurate, personalized models for them. We leverage two ideas to achieve this challenging goal: a) learning a compact interpretable latent representation for our task, and b) meta-learning the algorithm to effectively train person-specific networks in the few-shot manner without overfitting.
Watch this session
Join in the conversation below.