Bio-Plausible Back-propagation-free Learning for Edge Applications
Project Description
Future edge devices must continuously learn and adapt to new users and environments while operating under strict energy and memory constraints. However, backpropagation, the foundational algorithm for model training, conflicts with these constraints. It stores intermediate activations, propagates a global error backward, and updates weights across the entire model. Since activation storage far exceeds on-chip memory, this process generates off-chip data movement that dominates energy. In contrast, biological neurons learn locally, with dendritic branches integrating inputs and adjusting connections without a network-wide error. This project seeks to establish foundations for backpropagation-free/local learning. The research will explore dendritic computation as a source of local learning signals, scalable learning rules for deep neural network deployment at edge, and dendrocentric hardware architectures to support new algorithms for model training. This work has the broader goal of enabling energy-efficient, private, and adaptive intelligence on edge devices.
Start Date
January 1, 2027
Postdoc Qualifications
Someone who completed/completing their PhD by December, 2026. Has published in reasonable venues.
Co-advisors
Kaushik Roy, ECE, Purdue, kaushik@purdue.edu
Krishna Jayant, Biomedical engineering, Purdue, kjayant@purdue.edu
Bibliography
MPE Apolinario, A Roy, K Roy, "LLS: local learning rule for deep neural networks inspired by neural activity synchronization," 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV).
A Roy, MP Apolinario, S Das Biswas, K Roy, "Feedback Alignment Meets Low-Rank Manifolds: A Structured Recipe for Local Learning," The IEEE/CVF Winter Conference on Applications of Computer Vision 2026.
K. Roy, A. Jaiswal, P. Panda, "Towards spike-based machine intelligence with neuromorphic computing," Nature, 2019.
M Habibi Matin, S Xiao, K Jayant, "Mild Focal Cooling Decouples Dendrites to Reconfigure Cortical Output," Advanced Science, e20773.
M. P. E. Apolinario, K. Roy, C. Frenkel, "TESS: A scalable temporally and spatially local learning rule for spiking neural networks," IJCNN, 2025.