Neuroscience Grapples With How to Compare Brains and AI Models

Neuroscience has always relied on comparison. Darwin used it to build evolution. Kidney researchers in the 1960s compared the thickness of the renal medulla across desert and non-desert mammals and found a tight correlation with the animals' ability to concentrate urine—a clue that unlocked understanding of how countercurrent multiplication actually works. Comparison, done well, moves beyond mere cataloging. It builds mechanism. It explains why things work the way they do.

Neuroscientists have long compared brains too. We use anatomical atlases to identify matching structures across species. We note that hippocampus size correlates with spatial navigation ability. But until recently, we lacked the tools to do something harder: record from hundreds or thousands of neurons simultaneously and ask whether two brains—or a brain and an artificial network—are computing in the same way. The technology has arrived. Recording devices are cheaper, smaller, more standardized. We can now capture neural activity at a scale that was impossible a decade ago. And we have built artificial intelligence systems that, at least on the surface, resemble biological brains: distributed computation across large populations of simple units.

But we have not solved the fundamental problem. When you record from the same brain region in two different animals, how do you know if their neural responses are truly alike? When you compare a biological brain to an AI model, what does similarity even mean? The field has generated more than thirty different methods to answer these questions, and they do not agree. Some methods, like representational similarity analysis and centered kernel alignment, treat the problem geometrically—asking whether two systems arrange their responses in the same shape. Others favor prediction, measuring similarity by how well one system's activity can forecast the other's. Still others blend both approaches. A practitioner trying to navigate this landscape faces a bewildering choice, and many simply do not have time to understand the mathematical nuances of each approach.

The proliferation is partly an illusion. Dig beneath the surface and surprising patterns emerge. Representational similarity analysis and centered kernel alignment, routinely treated as separate tools, are formally equivalent once you add a mean-centering step to the first. Many other methods collapse onto a few underlying mathematical objects. Understanding this hidden unity helps. But it also reveals a deeper confusion: the field has conflated different kinds of questions. A predictive score tells you whether one system carries the information needed to reconstruct another. A geometric score tells you whether two systems organize information the same way. These are not the same thing. A neural network might predict brain activity perfectly while organizing that information in a completely different way. Conflating prediction with similarity invites error.

There is also the question of what counts as a proper metric. Many similarity measures are merely scores—numbers without mathematical structure. True metrics are symmetric and obey the triangle inequality: if system A is close to system B, and system B is close to system C, then A and C cannot be far apart. This sounds pedantic, but it is the difference between a number and a map. When a measure is a true metric, you can embed brains and networks into a common space, cluster them, and apply standard machine-learning tools. You can navigate coherently. You can ask new questions.

The deepest challenge, though, may be the simplest to state: brains are complex. No single metric will capture similarity across all experiments or between a model and a biological recording. The field should report multiple metrics, each capturing different aspects of neural computation. This requires hard work—digging into mathematical details and assumptions. But the reward is real understanding, not just rankings. The kidney researchers did not care about medullary thickness for its own sake. They cared because it pointed toward mechanism. Neural similarity scores matter only insofar as they illuminate how computation actually happens. The field has grown comfortable with leaderboards and novel metrics that are technically different but scientifically marginal. Both habits venerate the scores themselves over understanding. That needs to change. The tools exist. The question now is whether neuroscience will use them to build genuine insight.