If you are working on generative AI with an information theoretic approach, you may be interested in taking a look at our recent work on:
- Mutual information estimation – NeurIPS 2024
- Optimal Bayesian classification – ICML 2024
Mutual information estimation has many applications, among which is the evaluation of the ability of a communication system to deliver information from senders to recipients reliably.
Classification is another relevant task in AI. In communications, it coincides with the receiver’s task of decoding the received information. We exploited a variational representation of the f-divergence to derive a general class of value functions and solve the two problems in an optimal manner. Additionally, we proposed a new architectural approach based on derangements to improve training.
