Misalignment Between Backpropagation and the Hierarchy of Brain Responses to Images

Researchers using fMRI and MEG data found that while backpropagated gradients in deep neural networks can predict brain activity in higher visual cortex, their spatial and temporal organization fundamentally diverges from how the human brain processes visual information. This suggests that although artificial and biological neural networks may learn similar representations, they employ distinctly different learning mechanisms.

This neuroscience study challenges a prominent assumption in AI research: that backpropagation—the algorithm powering modern deep learning—mirrors how the brain learns. While previous research demonstrated that forward activations in neural networks correlate with cortical hierarchies, this work specifically interrogates whether backpropagation itself follows biological plausibility constraints. The researchers extended encoding analyses to map gradient flows from eight vision models onto neuroimaging data, revealing a striking disconnect. Backpropagated gradients do correlate with brain signals, but their computational flow and spatial arrangement deviate substantially from the brain's hierarchical organization. The findings suggest the brain either implements radically different learning algorithms or achieves similar representational capacity through alternative mechanisms entirely. This distinction matters significantly for AI development. If backpropagation isn't biologically implemented, researchers pursuing brain-inspired AI architectures may be pursuing a false parallel. The divergence could explain why scaling backpropagation hits certain efficiency ceilings where biological systems excel. For practitioners, this implies that borrowing more directly from neuroscience—rather than assuming biological plausibility in gradient-based learning—might unlock efficiency gains. The work also strengthens growing arguments for exploring alternative learning rules in artificial systems, potentially accelerating research into local learning, predictive coding, or other frameworks gaining traction in computational neuroscience.

• →Backpropagated gradients can predict neural activity in higher visual cortex but with fundamentally different spatial-temporal organization than brain hierarchies
• →Forward activations in neural networks align with visual processing hierarchies, but backpropagation does not follow the same biological constraints
• →The brain likely employs distinct learning mechanisms compared to artificial backpropagation despite achieving similar representational content
• →Results suggest brain-inspired AI architectures should move beyond assuming backpropagation is biologically plausible
• →Findings open pathways for exploring alternative learning algorithms that may better mirror biological neural computation