Sharpen Focus: Learning with Attention Separability and Consistency
This addresses the need for discriminative attention in image classification to reduce visual confusion, representing an incremental improvement over existing gradient-based attention methods.
The paper tackles the problem of overlapping attention maps causing visual confusion in convolutional neural networks by introducing a framework with learning objectives for attention separability and consistency, resulting in improved classification accuracy across multiple benchmarks, such as gains of +3.33% on CIFAR-100 and +5.73% on PASCAL VOC2012.
Recent developments in gradient-based attention modeling have seen attention maps emerge as a powerful tool for interpreting convolutional neural networks. Despite good localization for an individual class of interest, these techniques produce attention maps with substantially overlapping responses among different classes, leading to the problem of visual confusion and the need for discriminative attention. In this paper, we address this problem by means of a new framework that makes class-discriminative attention a principled part of the learning process. Our key innovations include new learning objectives for attention separability and cross-layer consistency, which result in improved attention discriminability and reduced visual confusion. Extensive experiments on image classification benchmarks show the effectiveness of our approach in terms of improved classification accuracy, including CIFAR-100 (+3.33%), Caltech-256 (+1.64%), ILSVRC2012 (+0.92%), CUB-200-2011 (+4.8%) and PASCAL VOC2012 (+5.73%).