Finding Concept Representations in Neural Networks with Self-Organizing Maps
This work addresses the interpretability of neural networks for researchers, though it is incremental as it builds on existing methods for concept representation.
The paper tackles the problem of identifying and locating neural representations of abstract concepts in neural networks by using self-organizing maps to inspect activation vectors, showing that relative entropy is a suitable measure for this purpose.
In sufficiently complex tasks, it is expected that as a side effect of learning to solve a problem, a neural network will learn relevant abstractions of the representation of that problem. This has been confirmed in particular in machine vision where a number of works showed that correlations could be found between the activations of specific units (neurons) in a neural network and the visual concepts (textures, colors, objects) present in the image. Here, we explore the use of self-organizing maps as a way to both visually and computationally inspect how activation vectors of whole layers of neural networks correspond to neural representations of abstract concepts such as `female person' or `realist painter'. We experiment with multiple measures applied to those maps to assess the level of representation of a concept in a network's layer. We show that, among the measures tested, the relative entropy of the activation map for a concept compared to the map for the whole data is a suitable candidate and can be used as part of a methodology to identify and locate the neural representation of a concept, visualize it, and understand its importance in solving the prediction task at hand.