A new study challenges the assumption that training improves how well neural networks mimic the brain. Researchers found that a single round of supervised training can reduce a model's alignment with the early visual cortex by 25–90%, depending on the learning rule used.
The Research
Nils Leutenegger, a researcher at the intersection of machine learning and neuroscience, trained neural networks on 720 object images from the THINGS database. He measured how well the models' internal representations matched human fMRI data from three subjects across six visual brain regions. The study, posted on arXiv in May 2026, tracked alignment at eight training checkpoints using representational similarity analysis (RSA).
Four learning rules were compared: backpropagation (BP), feedback alignment (FA), predictive coding (PC), and spike-timing-dependent plasticity (STDP). Results showed that a single epoch of training reduced V1 alignment by 25-90%. Backpropagation caused the largest drop in V1 alignment (Δr = -0.08), while predictive coding and STDP preserved more brain-like structure (Δr ≈ -0.04). Interestingly, the opposite trend appeared in object-selective cortex (LOC), where BP showed the largest increase in alignment, though the absolute change was small.
These findings suggest that untrained networks already capture low-level visual statistics through their architecture alone. Global error signals like backpropagation aggressively reshape early representations, while local learning rules (PC, STDP) better preserve brain-like structure.
Why It Matters
This research has important implications for both AI and neuroscience. It suggests that training neural networks with backpropagation may actually make them less brain-like, especially in early visual areas. The fact that untrained networks align well with the brain highlights the power of inductive biases built into architectures. For brain training platforms like IQGenio, this underscores that not all learning methods are equal—some may harness the brain's natural processing more effectively.
What You Can Do
While these findings are about artificial networks, they remind us that learning strategies matter. To keep your own visual system sharp, engage in tasks that involve local, feedback-driven learning—like puzzles that require spatial reasoning or pattern recognition. Avoid over-relying on rote memorization; instead, focus on understanding underlying structures.
Source: arXiv q-bio.NC
Curious about your own brain? Take our free adaptive IQ test or try 306 brain training levels.
