Rewriting the Rules of a Classifier

Abstract

Observations of various deep neural network architectures indicate that deep networks may be spontaneously learning representations of concepts with semantic meaning, and encoding a relational structure or rule between these concepts. We refer to these encoded relationships between concepts in the network as rules. In classifiers, we rewrite an existing rule in the network as desired, referred to as the rewriting technique.

We demonstrate that using our rewriting technique and simple human knowledge about how to classify the world around us, we can generalize existing classes to unseen variants, identify spurious correlations present in the dataset, mitigate the effects of spurious correlations, and introduce new classes. We find that our technique reduces the need for: computing resources, because we only re-train a single layer’s weights; new training images, because our rewriting technique can rewrite using concepts already encoded in the network; and domain knowledge, because what we choose to edit to improve classification is derived from logical rules a human would construct to classify images.