Semantic Bottlenecks: Quantifying and Improving Inspectability of Deep Representations

摘要

Today’s deep learning systems deliver high performance based on end-to-end training but are notoriously hard to inspect. We argue that there are at least two reasons making inspectability challenging: (i) representations are distributed across hundreds of channels and (ii) a unifying metric quantifying inspectability is lacking. In this paper, we address both issues by proposing Semantic Bottlenecks (SB), which can be integrated into pretrained networks, to align channel outputs with individual visual concepts and introduce the model agnostic Area Under inspectability Curve (AUiC) metric to measure the alignment. We present a case study on semantic segmentation to demonstrate that SBs improve the AUiC up to six-fold over regular network outputs. We explore two types of SB-layers in this work. First, concept-supervised SB-layers (SSB), which offer inspectability w.r.t. predefined concepts that the model is demanded to rely on. And second, unsupervised SBs (USB), which offer equally strong AUiC improvements by restricting distributedness of representations across channels. Importantly, for both SB types, we can recover state of the art segmentation performance across two different models despite a drastic dimensionality reduction from 1000s of non aligned channels to 10s of semantics-aligned channels that all downstream results are based on.