← SIGNALS
[TECH]

Sakana AI Develops Error Diffusion Technique for Training Dual-Stream Networks

Sakana AI has introduced a novel training method called Error Diffusion, which allows dual-stream networks to achieve high accuracy on MNIST and CIFAR-10 datasets without using backpropagation.

Editorial StaffJuly 18, 20261 MIN READ
Sakana AI Develops Error Diffusion Technique for Training Dual-Stream Networks

Sakana AI has made significant strides in the field of artificial intelligence with its new training method known as Error Diffusion. This innovative approach enables the training of dual-stream excitatory and inhibitory networks while adhering to Dale's principle.

The results are impressive, with the model achieving 96.7% accuracy on the MNIST dataset and 61.7% on the CIFAR-10 dataset. These benchmarks highlight the effectiveness of the Error Diffusion method in bypassing the limitations of traditional backpropagation.

By sidestepping the reliance on weight transport, which biological circuits may not be able to implement, Sakana AI's technique opens new avenues for developing more efficient neural networks.