Sakana AI's MNIST-to-CIFAR flip exposes a blind spot in brain-like AI

For years, attempts to train neural networks like brains work on MNIST, the handwritten-digit dataset from the 1990s, then fail on anything harder. Sakana AI has now pushed one such method, Error Diffusion, onto convolutional networks, CIFAR-10 images, and continuous-control reinforcement learning. But the important result isn't the accuracy. It's what happened when the team switched benchmarks. The company's work on routing problems to the right method suggests they think hard about benchmark specificity.

Why obeying the brain is expensive

Biological neurons obey Dale's principle: each neuron is either excitatory or inhibitory across all its synapses, never both. Standard artificial networks ignore this and let any connection be positive or negative. Backpropagation exploits that freedom by sending exact transposed copies of forward weights backward, a process neuroscientists call "weight transport" with no biological equivalent. Alternatives like Feedback Alignment, Direct Feedback Alignment (DFA), and predictive coding still allow mixed weights. Sakana AI's dual-stream Error Diffusion goes further: it splits each layer into positive and negative populations, keeps all four weight matrices per layer non-negative, and hardcodes the inhibitory sign into the architecture. That structural honesty costs about 4x the parameters of an unconstrained network of the same width (32 million versus 8 million for DFA).

The scores, in context

With a "modulo error routing" trick that assigns each hidden unit a fixed output channel, the method scores 96.7% on MNIST and 61.7% on CIFAR-10, the first time Error Diffusion has worked on convolutional networks at all. Compared to alternatives, the results are a trade-off, not a win. This kind of nuanced leaderboard reading echoes other work warning that benchmark numbers mask real-world performance.