Brain imaging reveals learning signals in human cortex don't match backpropagation

A team from Meta AI and France's CNRS compared gradient flows in modern vision models against high-resolution fMRI and MEG recordings of human brain activity and found no evidence that the brain uses a biological analog of backpropagation. The preprint on arXiv shows forward-pass representations in neural networks match human visual cortex activity closely, but the learning signals—the gradients that drive weight updates—follow a completely different pattern.

The researchers fed identical image sets to both trained vision models and human subjects, then mapped where feedforward activations and backprop gradients aligned with brain signals. Forward activations tracked the visual hierarchy from early to late cortical areas as expected. Gradients, however, showed no such correspondence. The mismatch held across multiple architectures and across both fMRI (spatial resolution) and MEG (temporal resolution) modalities.

How the brain might learn instead

The finding suggests the brain relies on a learning rule that artificial networks have yet to replicate. Backpropagation remains the dominant training algorithm in deep learning because it works, but it requires global error signals propagated backward through every layer—a mechanism with no clear biological substrate. The study adds to a growing body of work indicating the brain's training process is local, asynchronous, or both, and may be orders of magnitude more sample-efficient than gradient descent.

The paper does not propose an alternative algorithm. It rules out one hypothesis—that the brain secretly implements backprop under the hood—and leaves the door open for theories involving local Hebbian updates, predictive coding, or other biologically plausible learning rules. For practitioners, the implication is that the next leap in AI efficiency may come from reverse-engineering the brain's actual learning mechanism rather than scaling existing architectures.