Representational Similarity Analysis

Representational Similarity Analysis (RSA) is an approach that could help us understand the internal representation of deep learning neural network.

Why "Representational"

• representation matters: neural networks should be able to learn representations of latent variables without human labelled data
• invariance matters: representations should be invariant - that is what helps supervised learning
  

RSA can determine similarity

• RSA is a way of determining how similar the representations of different stimuli are to each other
• If a network has invariant stimuli, then RSA should show that objects from the same category are represented in a similar manner regardless of how they are transformed (e.g. orientation, location, etc.
• Representational similarity matrix