Saccader: Improving Accuracy of Hard Attention Models for Vision

TL;DR

Saccader is a novel hard attention model that improves interpretability and accuracy in image classification by using a pretraining step to guide attention, achieving competitive results on ImageNet.

Contribution

Introduces Saccader, a hard attention model with a pretraining step that enhances training efficiency and accuracy on complex datasets.

Findings

Achieves 75% top-1 accuracy on ImageNet

Attends to less than one-third of images

Narrowing the gap to baseline models

Abstract

Although deep convolutional neural networks achieve state-of-the-art performance across nearly all image classification tasks, their decisions are difficult to interpret. One approach that offers some level of interpretability by design is \textit{hard attention}, which uses only relevant portions of the image. However, training hard attention models with only class label supervision is challenging, and hard attention has proved difficult to scale to complex datasets. Here, we propose a novel hard attention model, which we term Saccader. Key to Saccader is a pretraining step that requires only class labels and provides initial attention locations for policy gradient optimization. Our best models narrow the gap to common ImageNet baselines, achieving $75\%$ top-1 and $91\%$ top-5 while attending to less than one-third of the image.