DeepBrain at Berkeley · fall 2026 · open study
Scrambled Brains
How accurate must a wiring map be before a brain built from it keeps its behaviour? We run a whole fruit-fly nervous system from its connectome, teach it to ride a bicycle, then corrupt the wiring step by step, from realistic mapping errors to random wiring, and measure where the riding breaks.
- 166,700 neurons, brain and nerve cord
- 25.6 M connections, 124 M synapses
- wiring ladder R0 to R5
- one task the bike cannot do by itself
The question
Whole-brain emulation assumes that a scanned wiring map plus a neuron model reproduces behaviour. Proofreading the map is projected to dominate the cost of future connectomes, yet nobody has measured how much wiring accuracy a closed-loop behaviour actually needs.
The sharpest objection is the "digital sphinx": Brunton et al. (eLife 2026) trained a decoder that let a 302-neuron worm connectome drive a walking fly body, and argued that a random network would have served as well. If a trained decoder can extract behaviour from scrambled wiring, closed-loop behaviour is not a valid test of an emulation at all.
So we grade the fly's wiring from real to random and score every rung on identical tasks: riding a bicycle below the speed at which it balances itself, the model's published decoder-free predictions, and how many neurons a decoder must read. The result is a fidelity curve: an acceptance test any connectome or emulator can be run against.
The wiring ladder
Every rung keeps the neurons and each connection's sender, so out-degrees and transmitter signs stay where they are. What else survives shrinks rung by rung. Generators and tests: fly-brain, brain/variants.py.
| Rung | What survives of the real wiring | Generator | Transmission pre-test | Riding |
|---|
Results so far
Two things are settled before any scrambled brain is raced. First, the real wiring can balance the bike below its self-stable speed, so the task is not solved by the bicycle. Second, the pre-test tells us which rungs still carry a left-right signal at all.
1. The balance gate: the real wiring learns to ride
Mean riding time per generation, 48 riders × 15 s, blank decoder at the start
Below the self-stable range (4.70 to 7.74 m/s); 5 N·m gusts; leaving the road ends a run. Dashed grey: the bike with no steering at all. Jobs 37491 to 37494, each stopped at its 5 h cap.
From a blank decoder, generation 0 rides 0.75 to 1.0 s on average, below the hands-off bike. After 57 to 63 generations the riders stay up 9.6 to 11.8 s at 4.0 m/s and 11.9 to 13.7 s at 4.5 m/s, with 23 to 77% riding the full 15 s. The hands-off bike manages 3.5 s (4.0 m/s) and 5.7 s (4.5 m/s) and never finishes. Two seeds per speed.
Table view
2. The transmission pre-test: do scrambled brains still carry a left-right signal?
Before any rung gets search compute, we hold the bike upright, drive every sense at cruise rate, and add 30 Hz to the left or the right vertical-system (VS) or haltere inputs in 8 brains. A rung passes when descending and motor neurons respond differently to left and right (|z| > 3; chance under 1) at a network rate within 2× of the real brain's. The synapse scale multiplies Shiu's weight; 0.47 is the male CNS calibration.
At the calibrated scale the two bottom rungs are almost silent, about 1% of the real network's activity, with at most one lateralised output. Degree-preserving wiring (R4) needs three times the synaptic weight to reach half the real activity and 14 to 17 lateralised outputs, spread over five times as many neurons. Random wiring (R5) stays at chance everywhere. The cause is convergence: neurons receiving 200 or more synapses from the bike's 789 sensory neurons fall from 417 to 26. The four new rungs (run 2 Nov on CPU) sharpen this: count jitter (R2) is indistinguishable from the real wiring, and cell-type-level wiring (R3) still carries the left-right channel at 74% of the real rate. But R3.5, which gives every neuron the right number of inputs, the right input strength and the right excitatory/inhibitory mix from the wrong partners, is as silent as R4. The transition lies between R3 and R3.5, and it is about partner identity, not input statistics. Transmitter errors (R1) halve the network rate and keep the haltere channel; the VS channel needs the 30-brain confirmatory run. One rewiring draw, 8 brains per cell; treat as a pilot.
3. The rungs generated this week
| Rung (seed 1) | Real connections kept | What the generator reports |
|---|---|---|
| R1 | 100% | 9,907 neurons relabelled (5.9%), 8,584 change the sign of their outputs (5.1%); inhibitory share of synapses 38.4% → 40.2% |
| R2 | 100% | Jitter factors with median 1.00 and a 16–84% range of 0.80–1.25; total synapses +2.6% |
| R3 | 39.4% | 25,267 (type, side) groups, 13,611 with a single neuron; 8.2% of neurons keep their inputs. Many groups are tiny, so a redrawn target often lands on the original neuron |
| R3.5 | 0.6% | 1,358 (sign, synapse-count) strata, every connection could move; in-degree, input strength and E/I input mix exact for every neuron |
| R4 | 0.58% | 0.34% of connections merged as duplicates |
| R5 | 0.27% | 0.13% merged |
Riding runs on R1 to R3.5 are queued behind the GPU node; their pre-test at the calibrated scale is in the table above. Every number above comes with its command and log in fly-brain's CURRENT_NUMBERS.md and JOBS.md.
Three ways to score a rung
Track A · closed-loop riding
A decoder learned from blank weights (cross-entropy method, identical seeds for every rung) reads the k descending neurons that the brain's own passenger screen ranks highest, k = 8 and 32, with a random-k control. Reported as the gap to two reference lines: a decoder that bypasses the brain and reads the senses directly (ceiling), and the hands-off bike (floor).
Track B · decoder-free predictions
Shiu et al.'s published feeding and grooming predictions and a handful of flybench tasks, rerun on every rung with no trained readout. Adds transmitter errors at Eckstein et al.'s measured confusion rates as a control the field does not yet have.
Track C · decodability
How many recorded neurons a decoder needs to reach 90% of the full roll-rate R² (N90), at a brain descending-neuron site and a nerve-cord motor site; what survives calcium imaging blur and slow frame rates; how fast decoding degrades when a probe misses its target by tens of microns.
What we predict
Preregistered before the ladder is raced (draft): R1 and R2 ride within one seed-SD of the real wiring. R3 loses under 30%, and the transition lies between R3 and the strength-preserving R3.5. R4, recalibrated to its own synaptic strength, rides no better than the sensory bypass. If R4 rides as well as R0, a trained decoder can extract the behaviour from scrambled wiring, closed-loop behaviour is not a valid test of an emulation, and we release the decoder-free battery as the test the field should use instead. Negative results get published as prominently as positive ones.
Milestones
Team
Led by Victor Chun (simulation, Track A) with DeepBrain at Berkeley students: a ladder engineer (generators, invariant tests, transmission gate), an error-budget lead (Track B), a decoding and statistics analyst (Track C, paired bootstrap) and an imaging and electrode lead (simulated calcium view, probe tolerance). Every member owns one figure panel and one section of the report. Agents generate rungs, run cluster jobs and draft reports; a human checks every number.