Revisit Recurrent Attention Model from an Active Sampling Perspective

Jialin Lu

TL;DR: Inspired by neuroscience research, solve three key weakness of the widely-cited recurrent attention model by simply adding two terms on the objective function.

Abstract:

We revisit the Recurrent Attention Model (RAM, Mnih et al. (2014)), a recurrent neural network for visual attention, from an active information sampling perspective.

We borrow ideas from neuroscience research on the role of active information sampling in the context of visual attention and gaze (Gottlieb, 2018), where the author suggested three types of motives for active information sampling strategies. We find the original RAM model only implements one of them.

We identify three key weakness of the original RAM and provide a simple solution by adding two extra terms on the objective function. The modified RAM 1) achieves faster convergence, 2) allows dynamic decision making per sample without loss of accuracy, and 3) generalizes much better on longer sequence of glimpses which is not trained for, compared with the original RAM.