Challenging the Assumptions: Untrained Neural Networks vs. Human Brain Alignment

In the ongoing quest to understand how artificial neural networks align with our own brain's visual processing, recent findings throw some conventional wisdom into disarray. It turns out that random, untrained neural networks are more than just noise, they can actually mirror the early visual cortex's representational patterns, often matching or even surpassing their trained counterparts.

Reevaluating Learning's Role

This revelation comes from an in-depth examination of representational similarity analysis (RSA) alignment with human fMRI data. Researchers compared the outcomes across four learning rules: backpropagation (BP), feedback alignment (FA), predictive coding (PC), and spike-timing-dependent plasticity (STDP). Using 720 object images from the THINGS database and fMRI data from three subjects across six visual regions, they measured Spearman correlations between model and brain dissimilarity matrices at eight points during training, specifically from epochs 0 to 40.

Here's where it gets interesting: a single epoch of training can reduce alignment with the V1 area of the brain by a staggering 25-90%, depending on the learning rule. Backpropagation, often regarded as a gold standard in training, is actually the most disruptive, causing a delta r reduction of -0.080. Conversely, methods like predictive coding and STDP maintain a stronger resemblance to human data, with reductions around -0.04.

Inductive Biases and Their Impact

What they're not telling you is that these findings suggest untrained neural networks may inherently grasp low-level visual statistics through inductive biases, without the need for extensive training. This challenges the traditional narrative that learning is the cornerstone of aligning machine and human cognition. If mere randomness can fetch such results, what does it say about our current methodologies?

In the object-selective cortex (LOC), the dynamics shift slightly. Here, backpropagation shows the most significant increase in alignment during training. Yet, the absolute change remains minor, indicating a nuanced dance between training and representational fidelity.

A Call for Scrutiny

Let's apply some rigor here. The assumption that extensive training equates to better brain alignment doesn't survive scrutiny in this context. Instead, these results beckon a reevaluation of how we interpret and value different learning rules. Are global error signals like BP reshaping early representations too aggressively, inadvertently erasing structures our brains maintain? Perhaps local learning rules like PC and STDP, which better preserve such structures, deserve more attention.

To be fair, the implications of these findings stretch beyond academia, hinting at a need for innovative approaches in machine learning, particularly in applications involving cognitive simulations. As we stand on the brink of a new understanding, the real question is: are we ready to rethink the foundational principles of neural network training?