Training without backpropagation.
NoBackProp
Year
2026
Role
Machine Learning · Research
Overview
An empirical comparison of backpropagation against biologically plausible, non-backprop learning rules — Forward-Forward, Feedback Alignment, Target Propagation and weight-perturbation methods.
Companion to the review paper “Beyond Backpropagation”, this project tests how alternatives to backprop behave in practice, with the aim of low-power, sustainable and neuromorphic-friendly training.
Goal
Compare backprop and non-backprop methods empirically, with code, plots and a written comparison.
Context
Builds on the Beyond Backpropagation review paper and colloquium work on training networks without backprop.
