Learning to stop while learning to predict

ICML2020 9-25-2020

Motivation

• Task-imbalanced meta learning: different tasks need different numbers of gradient steps for adaptation

• Deep learning based algorithms usually have a fixed number of iterations in the architecture.

could we learn to stop automatically?

Overview

• predictive model $\mathcal{F}_\theta$ : transforms the input x to generate a path of states $x_1, x_2, ...,x_T$

• stop policy $\pi_\phi$ : sequentially observes the states $x_t$ and determines the probability of stop at layer $t$

• variational stop time distribution $q_\phi$ : stop time distribution induced by stopping policy $\pi_\phi$

Reference

• https://icml.cc/Conferences/2020/ScheduleMultitrack?event=6279

• https://icml.cc/media/Slides/icml/2020/virtual(no-parent)-14-17-00UTC-6279-learning_to_sto.pdf

• https://www.youtube.com/watch?v=3VqzhP44Eic