Predictive coding networks (PCnets) provide a biologically inspired framework for classification and generative tasks. However, their generative performance is limited when handling multi-modal class distributions, as the standard weight decay method struggles to differentiate between multiple modes within a class effectively. To address this issue, we propose two extensions: (1) integrating an auto-encoder (AE) to enhance latent representations and (2) employing a predictive coding Hopfield network (PCHN) to capture attractor dynamics. We evaluate these models on synthetic multi-Gaussian datasets and a subset of MNIST, revealing that the AE-enhanced PCnet can generate samples representing distinct clusters. Meanwhile, the PCHN variant independently discovers and maintains cluster structures without explicit latent supervision. Our findings highlight the potential of predictive coding as a sturdy framework for learning multi-modal data distributions in a biologically plausible manner.
Ganjidoost et al. (Mon,) studied this question.