Study: Deep Learning’s Backpropagation Misaligned with Brain’s Visual Hierarchy

A new paper from researchers at Meta AI and French neuroscience labs addresses a long-standing debate: does the brain implement backpropagation, the core algorithm behind deep learning? Using fMRI and MEG recordings of human brain responses to natural images, the team extended standard encoding analyses to map not just forward activations but also backpropagated gradients from eight pretrained vision models (including DINOv3) onto neural data. They found that backpropagated gradients could reliably predict brain activity in higher-level visual cortex and at later latencies, suggesting some correspondence.

However, the spatial and temporal organization of these gradients diverged from what a biologically plausible backpropagation mechanism would predict. Specifically, the order in which gradients are computed and their spatial arrangement did not follow the temporal and spatial hierarchies observed in the human visual cortex. This misalignment indicates that while deep networks and the brain may encode similar representations, they likely rely on fundamentally different learning processes. The study uses 9 figures and spans 13 pages, providing the first direct comparison of backward gradients to brain dynamics.