Backpropagation is the algorithm that powers deep learning, but a new study suggests the brain may not use it — even though the representations in artificial and biological vision systems look similar.
The Research
Led by Joséphine Raugel from Meta and collaborators across multiple institutions, the research team used fMRI and MEG recordings of human brain responses to natural images. They compared how well forward activations — the standard signal flow through neural networks — and backpropagated gradients (the error signals used for learning) predicted brain activity in eighteen vision models, including the self-supervised DINOv3.
They found that backpropagated gradients could reliably predict brain signals in higher-level visual cortex and for later processing latencies. However, the spatial and temporal organization of these gradients did not match the hierarchy of human visual processing. In particular, the order in which gradients are computed (from later layers back to earlier ones) and their spatial layout diverged from the temporal and spatial hierarchies observed in the brain. This suggests that while the final representations may align, the learning mechanisms driving them differ.
Why It Matters
For years, researchers have debated whether the brain implements backpropagation. This study provides strong evidence against it, at least for visual learning. The finding implies that deep neural networks are powerful tools for predicting brain activity, but they are not accurate models of how we learn. Understanding the brain's true learning rules could inspire more efficient, biologically plausible AI and shed light on cognitive strengths and challenges in humans.
What You Can Do
Your brain likely uses a different algorithm than AI — one that is more flexible, less data-hungry, and more energy-efficient. To keep your learning mechanisms sharp, engage in diverse cognitive tasks: puzzles, new skills, social interaction, and physical exercise. These activities promote neuroplasticity and may tap into the brain's unique learning processes.
Source: arXiv q-bio.NC
Curious about your own brain? Take our free adaptive IQ test or try 306 brain training levels.
