🧬 biology

Supervised Training Rapidly Degrades Early Visual Cortex Alignment Across Biologically Plausible Learning Rules

This study demonstrates that supervised training rapidly degrades the alignment between artificial neural networks and the human early visual cortex, with global error-based learning rules like backpropagation causing significantly more disruption than biologically plausible local rules such as predictive coding and STDP, suggesting that untrained architectures inherently capture low-level visual statistics through inductive biases.

Original authors: Nils Leutenegger

Published 2026-06-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nils Leutenegger

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The Big Surprise: "Blank Slates" Look More Like Brains Than "Trained" Ones

Imagine you have a brand-new, untrained computer brain (a neural network). It's just a pile of random connections, like a box of tangled headphones. Surprisingly, when you show it pictures, its "thoughts" (internal patterns) look a lot like how a human brain (specifically the part that sees simple shapes and lines, called V1) reacts to those same pictures.

Now, imagine you start training this computer brain to recognize objects (like cats, cars, or chairs). You'd expect that as it gets smarter at its job, it would start thinking more like a human.

The paper's shocking finding is the opposite: As soon as you start training the computer, it stops thinking like a human brain. In fact, the more you train it, the less it resembles the early part of the human visual system.

The Experiment: Four Different Ways to Learn

The researcher, Nils Leutenegger, wanted to see why this happens. He tested four different "learning rules" (methods for how the computer updates its brain). Think of these as four different teaching styles:

Backpropagation (BP): The standard, super-efficient method used in most AI today. It's like a strict teacher who knows the exact answer for every single student and sends a precise correction message all the way back to the beginning of the class.
Feedback Alignment (FA): A slightly more "biological" method where the correction messages are a bit messy and random, like a teacher shouting corrections down a hallway without knowing exactly who is listening.
Predictive Coding (PC): A method where the brain tries to guess what's coming next and only learns from the mistakes. It's like a student who only pays attention when their prediction is wrong.
STDP: A method based on timing, where connections strengthen only if two neurons fire at the exact same moment. It's like a local neighborhood rule: "If you and your neighbor do the same thing at the same time, we become friends."

What Happened? The "One-Second" Crash

The researcher watched these computers learn over 40 rounds (epochs) of training. Here is what he found:

1. The "Strict Teacher" (Backpropagation) is the Most Destructive
When the computer used the standard Backpropagation method, its resemblance to the human brain's early vision center (V1) crashed by 90% in just one single round of training.

The Analogy: Imagine a sculptor (the AI) starting with a block of marble that already looks vaguely like a human face (the random weights). The "Strict Teacher" tells the sculptor, "You need to make this look like a specific cat, not a human face." The sculptor immediately chisels away the human-like features to focus entirely on the cat. The result is a great cat statue, but it no longer looks like a human face at all.

2. The "Local Learners" (Predictive Coding & STDP) Hold On
The computers using the more biological, local learning rules (PC and STDP) also lost some resemblance to the human brain, but they lost much less (only about 25–30%).

The Analogy: These sculptors are told to make a cat, but they are only allowed to make small, local adjustments based on what they see right in front of them. They manage to make a cat, but they accidentally keep some of the original "human face" features of the marble block intact.

3. The Trade-Off: Early Vision vs. Object Recognition
There was a twist in the higher parts of the brain (the area that recognizes complex objects, called LOC).

The "Strict Teacher" (Backpropagation) destroyed the early vision similarity but actually improved the similarity in the object-recognition area.
The "Local Learners" kept the early vision similarity but failed to build strong object-recognition skills.
The Takeaway: It seems like a trade-off. To become really good at recognizing specific objects (like a cat), the AI has to sacrifice the general, human-like way of seeing simple lines and shapes. The "Strict Teacher" forces this trade-off aggressively; the "Local Learners" are too gentle to make the switch.

Why Does This Matter?

The paper suggests that the reason random, untrained computers look like human brains is not because they are "learning" to do so. It's because their basic structure (the way they are built) naturally captures the statistics of the visual world, just like our V1 does.

When we train them to do a specific task, we force them to change that natural structure. The more "global" and precise the learning rule is (like Backpropagation), the more aggressively it scrubs away the brain-like structure to make room for task-specific skills.

Summary in a Nutshell

Before Training: Random AI brains look surprisingly like human brains.
After Training: They stop looking like human brains.
The Culprit: The standard way we train AI (Backpropagation) is the fastest at destroying this "brain-like" quality.
The Lesson: Nature might not use the "Strict Teacher" method (Backpropagation) because it destroys the brain's natural structure too quickly. Instead, the brain might use "Local Learners" (like Predictive Coding) that preserve the brain's natural shape while still trying to learn.

Important Note: The paper does not claim this means AI is useless, nor does it suggest we should stop training AI. It simply points out a fundamental difference between how our brains learn and how current AI learns, specifically regarding how they handle the early stages of vision.

Technical Summary: Supervised Training Rapidly Degrades Early Visual Cortex Alignment Across Biologically Plausible Learning Rules

Problem Statement
While deep neural networks (DNNs) trained on visual tasks are known to develop representations correlating with primate visual cortex, a counter-intuitive finding suggests that untrained, randomly initialized networks often match or exceed trained networks in alignment with early visual areas, particularly V1. This challenges the assumption that learning inherently improves brain alignment. The central question addressed is whether learning improves or erodes this alignment, and how different learning rules—ranging from biologically implausible to plausible—affect these dynamics over the course of training.

Methodology
The study investigates the evolution of representational similarity between neural networks and human fMRI data across training.

Architecture: A shared convolutional architecture (three convolutional blocks followed by a fully connected layer) was used for all conditions to isolate the effects of learning rules.
Learning Rules: Four distinct rules were implemented:
1. Backpropagation (BP): Standard supervised training with exact gradient descent.
2. Feedback Alignment (FA): Uses fixed random feedback weights for the backward pass instead of transposed forward weights.
3. Predictive Coding (PC): Utilizes local prediction errors and iterative refinement.
4. Spike-Timing-Dependent Plasticity (STDP): Updates weights based on spike-timing correlations in Poisson spike trains.
Training Protocol: Models were trained for 40 epochs on a subset of CIFAR-10 (8,000 images). Activations were extracted at eight checkpoints (epochs 0, 1, 2, 5, 10, 20, 30, 40) across five random seeds.
Evaluation: Representational Similarity Analysis (RSA) was performed using 720 object images from the THINGS database. Model Representational Dissimilarity Matrices (RDMs) were compared against fMRI RDMs from three human subjects across six visual Regions of Interest (ROIs): V1, V2, V3, V4, LOC, and IT. Alignment was quantified via Spearman rank correlation.
Statistical Analysis: Paired, one-sided permutation tests were conducted across the five seeds. Due to the limited number of seeds ( $n=5$ ), the minimum attainable one-sided $p$ -value is $\approx 0.031$ .

Key Results

Universal Degradation in Early Visual Cortex: All learning rules resulted in a reduction of alignment with V1 after just one epoch of training. However, the magnitude of degradation varied significantly:
- Backpropagation (BP): Showed the most severe degradation, losing 90% of its initial V1 alignment after a single epoch ( $r$ dropped from 0.102 to 0.011).
- Feedback Alignment (FA): Showed an intermediate drop of 49%.
- Predictive Coding (PC) and STDP: Preserved substantially more alignment, losing only 25% and 31% respectively. By epoch 40, PC and STDP retained significantly higher V1 alignment than BP or FA.
Generalization Across Early Areas: The degradation pattern observed in V1 extended to V2 and V3, with BP consistently showing the largest drops. The effect attenuated in V4.
Opposing Trend in Object-Selective Cortex: In the Lateral Occipital Complex (LOC), a weak, opposite tendency was observed. BP showed the largest increase in alignment during training (from $r \approx 0$ to $r \approx 0.011$ ), whereas local rules (PC, STDP) showed negligible changes. The authors note these LOC changes are small in absolute terms and were not subjected to significance testing, treating them as a suggestive tendency.
Seed Variability: The differences between learning rules were systematic and consistent across seeds, rather than random fluctuations.

Significance and Claims
The paper claims that supervised training universally degrades the alignment of early visual cortex representations, but the severity of this degradation is systematically dependent on the learning rule.

Inductive Biases vs. Learning: The results suggest that untrained architectures capture low-level visual statistics inherent to their architecture (convolutional filters, pooling, normalization) rather than through learning. Training reshapes these representations toward task-relevant features, moving them away from the general-purpose statistics encoded by V1.
Global vs. Local Error Signals: The degree of degradation correlates with the "globality" of the error signal. Rules propagating precise, layer-specific gradients (BP) are more destructive to early visual structure than local learning rules (PC, STDP), which better preserve brain-like structure.
Representational Trade-off: The opposing trends in V1 (degradation) and LOC (slight increase for BP) suggest a trade-off. The global error signal in BP may simultaneously degrade early visual structure while building task-relevant object representations in higher layers. Conversely, local rules preserve V1 alignment but fail to develop strong LOC-aligned representations, suggesting local credit assignment alone may be insufficient for building hierarchical object representations.
Implications for Cortical Learning: The findings highlight a tension between task optimization and representational brain-likeness. The brain appears to maintain strong V1 representations while developing object selectivity in higher areas—a combination none of the tested rules achieved perfectly. This implies that biological learning mechanisms may require a more nuanced approach than any single rule tested here, potentially combining local preservation with hierarchical credit assignment.

Limitations Acknowledged
The authors note several constraints: the use of a simple architecture (3 conv + 1 FC) rather than deeper networks; the domain shift between CIFAR-10 training and THINGS evaluation; simplified approximations of biological STDP and PC; a small fMRI sample size (three subjects); and limited statistical resolution due to the five-seed design. Additionally, the absolute correlation values are small ( $r < 0.10$ ), leaving the practical significance for downstream neural processing as an open question.