Computational Neuroscience

In the [YOU] on AI Field Guide

The cycle that began with [YOU] on AI asks what it means to see our machines clearly. Computational neuroscience supplies a precise diagnostic instrument: the brain is a working example of the capabilities artificial systems still lack, and each gap between brain and machine marks a place where biology has discovered something the engineers have not. The brain learns from a fraction of the data, runs on twenty watts, learns new things without overwriting the old, and does all of this grounded in a body that moves through a world it must survive in. None of these properties are exhibited by current AI systems. Each one is a clue.

The two-way method of computational neuroscience also models how the cycle itself should read: not looking for a verdict (are machines intelligent or not?) but looking for what each system reveals about the other. Artificial networks, now good enough to compare meaningfully with the brain, have supplied neuroscience with existence proofs—demonstrations that connection-weight learning at sufficient scale is enough to produce sophisticated perception—while the brain has supplied the machines with a long list of unsolved engineering problems. This reciprocal illumination is the intellectual structure Sejnowski calls most exciting, and it is a structure that clarifies rather than settles the deep questions.

Origin

The founding conviction of computational neuroscience is easy to state and hard to fully accept: the brain is a kind of computer, and to understand it one must ask what it computes and how. This is not the claim that the brain is like a digital machine; Sejnowski has always insisted on the differences, which are vast. The brain is slow where silicon is fast, parallel where computers are largely sequential, analog and noisy and wet where chips are clean and discrete. But beneath those differences lies a shared subject: both brains and computers take in information, transform it, and produce behavior, and both can be analyzed in terms of the operations that transformation requires. The shared subject is computation over representations, and that abstraction makes a unified science of natural and artificial intelligence possible.

The field's institutional architecture was built in the 1980s. Sejnowski and Churchland's 1992 book gave it a manifesto. The journal Neural Computation, founded by Sejnowski in 1989, gave it a venue. The NeurIPS conference, which Sejnowski has long helped lead, gave it a gathering. For most of this period the comparison between brains and artificial networks was more aspirational than productive: the networks were too simple for the comparison to teach much about either. What changed was capability. When deep networks became sophisticated enough to match brain recordings statistically, the comparison acquired teeth, and the conversation Sejnowski spent his career trying to start finally began.

Key Ideas

Three levels of analysis. Borrowing David Marr's framework, computational neuroscience analyzes any cognitive system at three levels: the computational level (what problem is being solved?), the algorithmic level (what procedure solves it?), and the implementational level (what physical substrate runs the procedure?). The power of the framework is that questions at different levels can be studied independently, and discoveries at one level constrain hypotheses at the others. It also makes brain and machine commensurable: the same question can be asked of neurons and weights.

The existence proof. A central contribution of computational neuroscience has been demonstrating, by building working systems, what is in principle possible with learned connection weights. Before deep learning, it was genuinely unclear whether a network of simple units could produce sophisticated perception without additional ingredients. The success of deep networks at vision and speech settled a version of that question: at sufficient scale, the approach is enough. This is real information about what the brain might also be doing, even if it does not prove the brain does the same thing.

The remaining gaps as clues. The brain does things current AI systems cannot: it learns from far less data, runs on far less power, learns new things without overwriting old ones (catastrophic forgetting is a solved problem for biology), and grounds its knowledge in a body acting in a world. Computational neuroscience treats each gap as a research target, not an embarrassment. Each gap marks something the brain has discovered over evolutionary time that the engineers have not yet figured out—and therefore a place where studying the biological system can improve the artificial one.

The opacity parallel. A trained neural network is opaque: its knowledge is distributed across millions of weights, none of which means anything alone, and its competence cannot be reduced to rules a human could read. Sejnowski observes that the brain is opaque in exactly the same way. No neuroscientist can read knowledge off a synapse. From this vantage, the opacity of artificial systems is not a scandal but a family trait, shared with the only other systems known to be genuinely intelligent. The methods being developed to understand artificial networks—interpretability, mechanistic analysis, probing—are increasingly the same enterprise as the methods being developed to understand the brain.