B-cos Networks: Alignment is All We Need for Interpretability

TL;DR

This paper introduces B-cos networks, replacing linear transforms with B-cos transforms to enhance interpretability by aligning weights with task-relevant features, without sacrificing performance.

Contribution

The paper proposes B-cos transforms for deep neural networks, enabling highly interpretable models that align weights with features while maintaining standard architecture performance.

Findings

B-cos transforms induce interpretable linear summaries of the model.

Models with B-cos transforms perform comparably to standard models on ImageNet.

The approach is compatible with common architectures like ResNets and VGGs.

Abstract

We present a new direction for increasing the interpretability of deep neural networks (DNNs) by promoting weight-input alignment during training. For this, we propose to replace the linear transforms in DNNs by our B-cos transform. As we show, a sequence (network) of such transforms induces a single linear transform that faithfully summarises the full model computations. Moreover, the B-cos transform introduces alignment pressure on the weights during optimisation. As a result, those induced linear transforms become highly interpretable and align with task-relevant features. Importantly, the B-cos transform is designed to be compatible with existing architectures and we show that it can easily be integrated into common models such as VGGs, ResNets, InceptionNets, and DenseNets, whilst maintaining similar performance on ImageNet. The resulting explanations are of high visual quality and perform well under quantitative metrics for interpretability. Code available at https://www.github.com/moboehle/B-cos.