Search Papers | Poster Sessions | All Posters
Poster A96 in Poster Session A - Tuesday, August 6, 2024, 4:15 – 6:15 pm, Johnson Ice Rink
Bidirectional Predictive Coding
Gaspard Oliviers1 (), Mufeng Tang1, Rafal Bogacz1; 1University of Oxford - Nuffield Department of Clinical Neurosciences, Medical Research Council Brain Network Dynamics Unit
Predictive coding networks (PCNs) offer a compelling model for cortical sensory processing, yet their traditional unidirectional formulations—focusing solely on top-down or bottom-up processing—limit their applicability across various tasks and fail to capture the sensory system's bidirectional nature. We introduce a bidirectional predictive coding model, leveraging a single neural network for both top-down and bottom-up processing. This approach not only maintains the biological plausibility of PCNs but also excels in generative and discriminative tasks simultaneously within a single training session. Our model presents a unified, mechanistic view of how the sensory system employs a bidirectional approach to achieve its functionalities.
Keywords: Predictive coding Sensory processing Discriminative and generative tasks