Sakana AI trains convolutional networks without backpropagation, hitting 96.7% on MNIST
Sakana AI's July 17 arXiv paper extends its 'Error Diffusion' method to convolutional networks and reinforcement-learning agents for the first time, reaching 96.7% on MNIST and 61.7% on CIFAR-10 without backpropagation or arbitrary-sign weights. Because every weight stays non-negative, this is one of the first biologically faithful learners that can run natively on neuromorphic chips — photonic and analog hardware that standard deep learning cannot touch. It still trails the nearest backprop-free method by 7.4 points on CIFAR-10, which marks the current accuracy cost of dropping backpropagation.
Source: techtimes.com ↗
A learning rule that works with non-negative weights only is not just biologically faithful — it is directly deployable on a class of hardware that standard deep learning cannot touch.
Why this matters
- → First biologically faithful learner deployable on neuromorphic hardware (photonic, analog chips).
- → Eliminates weight-transport problem without random matrices—all weights non-negative.
- → Still trails backprop-free competitors by 7.4 points; quantifies accuracy cost of biological plausibility.