Functional correspondence without isomorphism: evaluating the brain-ANN analogy
A term essay: the brain does not run backpropagation, and language models still out-predict neuroscientists on their own literature. Marr's levels let both be true; the brain-ANN analogy lives at the computational level.
Written for a PSY290H1F (Behavioural Neuroscience) written assignment, submitted November 24, 2025. Rewritten here from the APA manuscript: the citations became inline links and the register is looser; the argument is unchanged. The manuscript, recompiled as submitted, is available as a PDF.
Large neural networks and the human brain now turn in strikingly similar behavioural performances on object recognition and language understanding, and the tempting conclusion, that they are the same kind of system, gets voiced regularly. Behavioural neuroscience has a precise word for the strong version of that claim: isomorphism, a structure-preserving mapping in which a model’s parts and operations correspond one-to-one with identifiable biological mechanisms. My position: the evidence rejects isomorphism at the implementation level, and still supports a restricted, genuinely explanatory correspondence at the computational level.
The learning rules do not match
Biological synaptic plasticity is local and activity-dependent. Hebbian learning and spike-timing-dependent plasticity update a synapse from the activity of the two neurons it connects, and decades of work on long-term potentiation and depression show neuromodulators like dopamine gating that plasticity by behavioural significance, not by any single scalar loss. Cortical circuits assign credit through distributed local rules, under constraints of biophysics and developmental history. No neuron sees a global error signal.
Deep networks are trained by backpropagation, which needs exactly what cortex does not have: weight copies shared between forward and backward passes, symmetric connectivity between layers, and a loss gradient propagated precisely, layer by layer. The interesting test of whether the brain could be doing something backpropagation-like comes from Shervani-Tabar and Rosenbaum, who meta-learned plasticity rules constrained to be biologically plausible, using fixed random feedback pathways instead of exact weight transport. The discovered rules train deep networks online and beat earlier random-feedback methods, yet still typically reach only 70 to 80 percent of backpropagation’s performance. Even optimized as hard as we know how, biologically constrained learning does not converge on gradient descent. The brain solves credit assignment some other way. On learning rules, the isomorphism claim is dead.
And yet the models predict neuroscience
The natural next step would be dismissing deep networks as engineering tricks with nothing to say about brains. The recent evidence refuses to cooperate. BrainBench is a forward-looking benchmark: given two versions of a neuroscience abstract, the original and one with a subtly altered result, decide which one reports what the experiment actually found. Across two hundred such cases, large language models scored about 81 percent; human neuroscience experts averaged about 63. BrainGPT, a Mistral-7B tuned further on the neuroscience literature, did better still, and the models’ confidence tracked their correctness, which suggests they had internalized real regularities about methods and plausible outcomes rather than guessing fluently.
Whatever is happening there, both the experts and the models face the same abstract problem: from partial descriptions of methods and prior findings, infer the most likely outcome. Predictive learning over the literature turns out to be sufficient for that task at expert level and beyond. And the parallel extends to how abilities arrive: emergent abilities in LLMs appear abruptly once scale and training cross thresholds, the way fluent language comprehension and fine-grained visual recognition emerge in development only after extended exposure during sensitive periods, a trajectory the course’s lectures on neurodevelopment and learning trace in detail. A system with none of the brain’s learning machinery still reproduces the shape of the brain’s learning curves.
Marr sorts the ledger
Marr’s three levels hold the two halves apart cleanly. The computational level asks what problem the system solves and under what constraints; the algorithmic level asks how information is represented and transformed; the implementational level asks what hardware runs it. Ku and colleagues argue that LLMs, like the cognitive models before them, should be compared to minds primarily at the computational level, with the lower levels treated far more cautiously.
Read the evidence through that frame and the ledger balances. The learning-rule mismatch lives at the algorithmic and implementational levels; that is where isomorphism fails, and the meta-learning results quantify by how much. BrainBench and the emergence parallels live at the computational level; that is where the correspondence is real, and where a model can earn explanatory standing the way Guest and Martin say any computational model must: by predicting, and by generating hypotheses someone can test. The course’s methods lectures add the caution that keeps this honest: predictive accuracy alone never proves shared mechanism. A model can be right for reasons the brain does not use, which is exactly what the learning-rule evidence says is happening.
The level it earns
The brain does not run backpropagation, and no amount of benchmark performance makes it so. But a model can tell us something true about the problems the brain solves while sharing none of its mechanisms. Large networks are working demonstrations that certain computational problems admit solutions at all, and of what those solutions require: integrating noisy findings into predictions, growing abrupt capabilities out of gradual experience. That is a real contribution, made at exactly one of Marr’s levels, and it survives every difference in wiring.