"The Brain Is a Neural Network"? Misra Clashes with NeuroAI Researchers
From September 26-28, Vishal Misra published a series of posts disputing the popular claim that "the human brain is itself a neural network, so AI will one day become like the brain." His core argument is that the word "literally" doesn't hold: modern AI's neural networks descend genealogically from a mathematical abstraction, not from biological neurons. The debate then drew in more participants—AI researcher arannayebi cited NeuroAI literature arguing that task-optimized ANNs correspond to brain regions; on September 28, coherence reposted and endorsed neuroscientist camhberg's rebuttal, noting that frontier AI models' representations can already predict primate cortical activity—but the two sides have not converged.
Confirmed
- Genealogy: all contemporary artificial neural networks (including Transformers) stem from one lineage—biological neurons → the 1943 McCulloch & Pitts mathematical abstraction → the perceptron → multilayer networks → deep learning.
- Visual comparison: Misra illustrated the gap with a side-by-side of Cajal's hand-drawn biological neuron and the McCulloch-Pitts mathematical abstraction.
- Complexity gap: modeling a single cortical neuron requires a neural network 5-8 layers deep (Beniaguev et al., 2021); an artificial neuron is a "cartoon" of its biological counterpart.
- Misra also noted that backpropagation has no biological counterpart.
- The other side: nayebi responded by citing a decade of NeuroAI literature showing that task-optimized ANNs best match brain regions (see footnote 6 of his paper), implying the analogy is scientifically grounded.
- Added September 28: coherence reposted and endorsed neuroscientist camhberg's rebuttal. That side acknowledges key differences between neural networks and the brain (substrate, neuromodulation, recurrent structure), and no one seriously claims machine intelligence equals the human brain; but they argue there is empirical grounding for discussing frontier AI systems' dynamics and representations—frontier models' representations can already predict primate cortical activity.
Unconfirmed
- The disagreement has not converged: Misra focuses on the failure of "literal mechanistic identity," while nayebi and camhberg emphasize empirically supported "functional correspondence." Whether these two lines of argument truly conflict remains to be seen.
Why it matters
The debate targets an oversimplified narrative dating back to Rosenblatt's perceptron: the term "neural network" invites people to assume AI and the brain share a mechanistic basis. Misra's evidence shows the complexity gap at the single-neuron level spans orders of magnitude, so "the brain is also a neural network" fails as a literal argument that AI will attain humanlike intelligence. Meanwhile, nayebi's and camhberg's responses represent NeuroAI's mainstream research program on functional ANN-brain correspondence—the progress in frontier-model representations predicting cortical activity turns the debate from conceptual clarification into an empirically weighted exchange. The dispute itself illustrates the tension between "analogy" and "identity" positions.
2026-09-26 ~ 2026-09-28 · 7 related posts
Primary sources
- Neural networks don't literally mirror the brain: the name descends from a 1943 math abstraction — vishalmisra ·
- Researchers clash: task-optimized ANNs match brain areas, but is the brain literally a neural network? — vishalmisra ·
- Neuroscientist pushes back: task-trained nets converge on brain-like representations that can predict primate cortex — coherence ·
- [source] Neural networks don't literally mirror the brain: the name descends from a 1943 math abstraction — vishalmisra · 2026-09-26
- Cajal's neurons vs. McCulloch & Pitts' abstraction: where AI neural nets actually come from — vishalmisra · 2026-09-26
- From Cajal to Transformers: the intellectual lineage of neural networks traced — vishalmisra · 2026-09-26
- One cortical neuron needs a 5-8 layer deep net to model, argues Columbia's Vishal Misra — vishalmisra · 2026-09-27
- One cortical neuron needs a 5-8 layer net to model — backprop has no biological analog — vishalmisra · 2026-09-27
- [source] Researchers clash: task-optimized ANNs match brain areas, but is the brain literally a neural network? — vishalmisra · 2026-09-27
- [source] Neuroscientist pushes back: task-trained nets converge on brain-like representations that can predict primate cortex — coherence · 2026-09-28