Training AI Without Backpropagation

Standard AI training relies on a process called backpropagation. This method requires data to move in ways that real biological circuits cannot perform. Researchers at Sakana AI wanted to change this by creating a system that follows biological rules.

They developed a technique called Error Diffusion. It allows networks to learn by splitting signals into two separate streams. This approach follows Dale's principle, which dictates how neurons in our brains connect to each other. By removing backpropagation, the model stays much closer to how humans learn.

The results show this method works well on standard testing tasks. The system reached 96.7 percent accuracy on simple number recognition and showed strong performance on more complex image data. This is a big step toward making artificial intelligence more efficient and biologically realistic.