Rethinking Learning: Beyond Backpropagation Toward Brain-Inspired Computational Intelligence

Abstract:
Backpropagation has powered the modern era of computational intelligence, enabling breakthroughs in perception, language, control, and autonomous systems. Yet as intelligent systems move into dynamic, real-world environments, new demands emerge: continual adaptation, robustness under uncertainty, energy efficiency, and scalable autonomy. These challenges invite a deeper question — are our learning algorithms fundamentally aligned with how intelligence itself operates?
This lecture explores predictive coding as a compelling, brain-inspired alternative for credit assignment in deep systems. Rather than relying on staged forward and backward passes with global error transport, predictive coding formulates learning as the continuous minimization of hierarchical prediction errors through local, parallel, and bidirectional interactions. Recent theoretical advances demonstrate that such dynamics can approximate gradient-based optimization, offering a principled bridge between neuroscience and modern machine learning.
This perspective reframes learning as an energy-minimizing dynamical process, opening new directions in distributed credit assignment, continual learning, robust inference, and neuromorphic implementation. By revisiting the principles of biological intelligence, this lecture argues that the next generation of computational intelligence systems may emerge not from scaling existing algorithms, but from rethinking the foundations of learning itself.
Co-sponsored by: IEEE VIC CIS Chapter; IEEE VIC Section
Speaker(s): Narayan Srinivasa,
Virtual: https://events.vtools.ieee.org/m/554620