In this paper, we propose a new interpretation of the information maximization method (InfoMax) from a perspective of the rate distortion theory. We show that under specific conditions, InfoMax is equivalent to the minimization of a compression rate under the constraint of zero distortion. Zero distortion, or equivalently, zero reconstruction error between the input and its reconstruction, does not provide meaningful solutions in many cases. Based on the new interpretation, we extend InfoMax to be able to deal with non-zero distortion and also to learn under/over-complete representations. Experimental results on synthetic as well as real data show the effectiveness of our method.
- Information maximization
- Rate distortion
- Unsupervised representation learning
ASJC Scopus subject areas
- Computer Science Applications
- Cognitive Neuroscience
- Artificial Intelligence