How well do artificial neural networks mirror the brain's visual system? A new preprint suggests the answer depends on which brain area and which species you're looking at. The study tested five learning rules against both human fMRI and macaque electrophysiology, finding that early visual areas are more aligned than higher-order areas—and that a pretrained ResNet-50 outperforms custom models in higher areas.
The Research
Nils Leutenegger, a cognitive researcher, compared five learning rules—backpropagation (BP), feedback alignment (FA), predictive coding (PC), spike-timing-dependent plasticity (STDP), and random weights—across species. Using representational similarity analysis (RSA), he measured alignment with macaque V1/V2 (102–103 neurons) and V4/IT (88–168 neurons) from two datasets, and compared those to human fMRI data from a companion study.
Key findings: all models showed higher alignment with macaque early visual cortex (V1/V2: rho = 0.15–0.30) than with human fMRI (rho = 0.01–0.08), likely due to the higher signal-to-noise ratio of electrophysiology. STDP and PC topped early visual alignment (rho ~ 0.30 and 0.28, respectively). However, at IT (inferior temporal cortex), rankings across species showed zero correlation (Kendall's tau = 0.00, p = 1.00), though the study had only five learning rules, limiting power to detect any but exact reversals. A pretrained ResNet-50 achieved rho = 0.25 at IT, beating all custom CNN conditions (rho = 0.07–0.14), suggesting that IT alignment is limited by model capacity and training data, not just the learning rule.
An August 2026 correction noted that a bug affected PC and STDP results; after repair, STDP still led in V1, but PC dropped to baseline. Other findings remained unchanged.
Why It Matters
This cross-species comparison is a rare step toward validating AI models of vision against biological brains. It shows that early visual processing is fairly conserved across species, which is good news for using AI to understand low-level vision. But the IT null result highlights that higher-order cognition is still poorly captured by current deep learning models—the gap between artificial and biological vision remains wide.
For readers, this underscores that AI, while impressive, still has a long way to go before it can mimic the brain's full complexity. It also demonstrates the value of studying multiple species to understand human cognition.
What You Can Do
- Challenge your visual processing: Try visual puzzles or brain-training games that target pattern recognition.
- Learn about cognitive science: Understand how your brain works by reading about research like this.
- Take a free IQ test: Measure your own cognitive abilities to see where you stand.
Source: arXiv q-bio.NC
Curious about your own brain? Take our free adaptive IQ test or try 306 brain training levels.