Bing Brunton on the Connectome, Neural Computation, and How the Brain Controls the Body

Computational neuroscientist Bing Brunton explains what a wiring map of the brain can and cannot tell us — using the newly completed fruit-fly connectome, a three-neuron walking circuit, and a pointed cautionary tale about what it means to simulate a brain.

Key ideas

1. A connectome is a wiring diagram, but the word is slippery and the map is partial. The connectome is the matrix of which neurons connect to which — sparse, structured, and asymmetric (a neuron can speak to another without listening back). But ‘connectome’ is used at wildly different scales — cell-by-cell for small animals, coarse-grained area-by-area for the human brain — and even the finest matrix omits what a neuron is (a dopamine cell and a serotonin cell firing the same spike say different things) and how its message is received. Half the cells in a brain are not even neurons.

2. Having a connectome is not understanding it. The 300-neuron wiring of the worm C. elegans has been mapped for over thirty years and is still not understood, because the worm also computes chemically and mechanically — squirting neurotransmitters and exploiting the physics of its own squishy body. The fruit fly’s larger map (about 150,000 neurons in the brain, 22,000 in the ventral nerve cord, published only within the past year) is paradoxically more tractable: it has genuine cell types, not one-off snowflake cells.

3. Simulate-then-prune found a walking rhythm in just three neurons. Brunton’s team simulated the fly’s leg-controlling circuit (about 4,000 cells) as a brute-force numerical model, then deleted neurons one at a time until only the minimal rhythm-generating core survived. The answer was three cells — two excitatory (E1, E2) and one inhibitory (I1) — a rare case where the model preceded the experiment and made a prediction later confirmed at the bench.

4. The brain does not live in a jar. Nervous systems always controlled a specific body with specific limbs and sensors, so Brunton wants to embody simulated connectomes inside physics-engine ‘digital twins’ of animals. Her digital Sphinx rebuttal shows why this matters: with enough deep learning, a worm connectome can drive a realistic fly body convincingly. Behaviour fidelity is no proof of biological fidelity — what makes a model meaningful is engineering the right interfaces, not matching the animation.

5. Cognition is built on sensorimotor control. The only agents we agree are intelligent and conscious are embodied ones. All nervous systems evolved to move a body through the world and respond to its senses; reasoning and awareness were built on top of that machinery. Understanding the sensorimotor substrate, Brunton argues, constrains how the higher functions could possibly have arisen.

Summary

The connectome and what it leaves out

A connectome is, roughly, the wiring diagram of a brain — an engineering schematic of which cells connect to which. Brunton is careful to say cells, not neurons: about half the cells in a brain are glia, long dismissed as custodial but now understood to have their own dynamics and vital functions. To a computationalist the fine-grained connectome is a giant connectivity matrix — sparse, highly non-random, and asymmetric, because a neuron can send to another without receiving back. But the word is used inconsistently: for large animals like humans the ‘connectome’ is coarse-grained to brain areas, not individual cells. And the matrix is only part of the story. The chemical identity of a cell matters (a dopamine spike and a serotonin spike carry different meanings), as does how the receiving cell is wired to interpret it — Brunton’s analogy is that the same sentence said to two people lands differently depending on the relationship. The finer biophysical parameters — ‘crazy non-linear’, a whole PhD to measure one cell — are almost certainly not all needed for a working model, but which of them are crucial is not yet known.

From worm to fruit fly: which wiring maps we have

The first connectome, mapped over thirty years ago, was the roughly 300 neurons of the nematode C. elegans — and it is still not understood. The reason, learned slowly, is that the worm does much of its computation off the wiring diagram: heavy chemical signalling and mechanical reflexes coupled to its soft, crawling body. The neural matrix alone was simply not enough. The breakthrough of the past year is the full connectome of the Drosophila fruit fly — about 150,000 neurons in the central brain plus 22,000 in the ventral nerve cord (the insect analogue of our spinal cord), mapped for one male and one female. Counter-intuitively, the bigger map is easier to interpret: the fly has jointed limbs like ours and enough cells to form real cell types, so the matrix is more directly informative than the worm’s compressed, multiplexed nervous system where single cells serve several senses at once.

Three neurons for a rhythm: the walking circuit

With John Tuthill’s experimental lab, Brunton attacked a century-old question: how does a nervous system generate rhythm from non-rhythm? Nearly all locomotion — walking, swimming, slithering, wing-flapping — is cyclic, as are breathing and digestion, driven by central pattern generators (CPGs). Roboticists happily model these as abstract oscillator equations, but nobody had ever named the actual cells that generate a walking rhythm in a real animal. Taking a reductionist bet — ‘it’s got to be in there somewhere’ — the team simulated the roughly 4,000 cells controlling two front legs (no machine learning, just numerical simulation of the connectivity matrix), got motor rhythms to emerge, then pruned neurons one by one until the rhythm broke. The minimal circuit was three cells: two excitatory (E1, E2), one inhibitory (I1), in an understandable architecture. Strikingly, the model made predictions before the experiments — one predicted that zapping a particular neck neuron with a laser should make the leg tap, which it did. The three cells generate the rhythm; many more are needed to turn that rhythm into coordinated walking.

The embodied brain and the digital-twin caution

Brunton’s larger vision is to put the connectome back inside a body — a biologically interpretable, physically realistic ‘video-game body’ in a physics engine, an agent that can walk on bumpy surfaces and interact with other agents. She borrows the term digital twin from industry (a simulation of, say, Singapore, fed by real-time sensors). But she is sharp about the failure mode. A startup claimed to have ‘uploaded a fly brain’; the claim went viral. To show why it was hollow, Brunton’s lab took the C. elegans worm connectome — all 300 cells — dropped it into a reinforcement-learning system, and had it drive a biomechanically realistic fly body that walked convincingly. If enough deep learning is doing the work and parameters are free to change, a worm brain in a fly body proves nothing: you get behaviour fidelity with zero biological fidelity. It is emulation — like running the same software on a Mac or a PC. What makes a model meaningful is engineering honest interfaces, not matching the camera footage.

What wiring maps explain — and therapeutic hopes

Pressed on consciousness, Brunton makes a deliberately strong claim: the only agents we all agree are intelligent and conscious are embodied ones, and every nervous system evolved — starting around 500 million years ago — to move a body and respond to its senses. Reasoning and awareness are built on that sensorimotor foundation, so understanding the substrate constrains the rest. She is more sceptical that biological detail is needed to build artificial intelligence — bird flight inspired aviation, but fixed-wing aircraft threw out the biology — though she suspects an intelligent agent may need to be embedded in something with physics, constraints, and conservation laws. The nearest therapeutic payoff is the interface between nervous system and musculoskeletal system, where spinal injury, amputation, and the slow, holistic adaptations of gait and plasticity are poorly understood; embodied models that can be ‘broken’ in controlled ways might yield new clues for rehabilitation.

Speakers

• Bing Brunton — computational neuroscientist; Professor of Biology, University of Washington. Studies how large-scale neural connectivity produces movement, and how simulated connectomes can be embodied in physical models.
• Sean Carroll — host; theoretical physicist and author, Mindscape.

See also

• Sean Carroll — host
• Consciousness — Brunton’s embodiment argument: the only agents we agree are conscious are embodied ones
• Anil Seth on Consciousness, Machine Sentience, and Why Intelligence Isn't Awareness — a related Mindscape-adjacent treatment of embodiment and awareness