A fixed fly connectome learns to remember: the cue hides in neuronal adaptation and is read only in the brain state of learning
Synopsis
A spiking model built from the complete central nervous system connectome of a male fly learns to plan moves and to remember a visual cue while its wiring stays fixed. The cue is not held by persistent firing but written into neuronal spike-frequency adaptation, kept silent through a blank delay and read out by the next input. Recall works only in the brain state of learning: the same trace is nearly reversed at an untrained delay, and a lesion mimicking the Kv4 loss caused by Aβ42 rewrites the trace and its reading, while learning under the lesion is spared.
Interpretation
A fly central nervous system model with completely fixed wiring can learn to plan moves and to remember a visual cue, and the memory needs neither a dedicated memory circuit nor persistent firing. Earlier connectome models only made predictions, or confined training to a single sensory pathway, and the connectome was often described as a static map without dynamics, state or plasticity. The model has 167,106 leaky integrate-and-fire neurons with spike-frequency adaptation and 25.6 million fixed connections; training changed only 35,310 gains and thresholds set per cell type plus the 2,052 weights of a linear readout, and the final brain with the final readout solved all 24 planning maps, whereas no other pairing of brain and readout checkpoints solved more than 3.
The cue is written into neuronal spike-frequency adaptation, which produces no differential spiking during the delay until the next input reads it out. In a single trial with a 640 ms delay, spikes that tell the two cues apart stop 22 to 42 ms after the cue ends, yet the cue persists. In one trial, 2,769 cells hold a difference in spike-frequency adaptation between the cues through the delay, and that difference decays with the adaptation time constant of 200 to 202 ms; in none of the four learners does a cell spike differently during the last 598 ms of the 640 ms delay.
Readout is locked to the timing and excitability state of learning, so changing state cuts access to an existing memory without changing storage itself. Net memory use is 1 at all 9 trained timings and averages −0.004 at the other timings, and the lock holds in 66 of 66 runs across nine scenes. After the untrained 320 ms delay, the decision-time pattern has a cosine of −0.70 with the trained 160 ms pattern and 92% of the strongest cells change sign; a lesion mimicking the Kv4 loss caused by Aβ42 lowers the trained histories solved by the learned readout from 18 to 7 at dose 1, while readouts trained in the lesioned brain solve 18 of 18 in every one of 48 runs.
Perspective
The conclusions come from a spiking model built on the connectome of one male fly central nervous system, and every statement about flies or patients is a prediction awaiting a test in the animal; the task delays all last under one second, shorter than the visual memories of several seconds in behaving flies.
The model has no neuromodulation, and synapses change only by short-term depression at olfactory terminals; the transplant between brain states moves the full state including its cue-independent part, so it separates where a trace was written from where it was read, not the trace from the state that carries it; the molecular links rest on transcripts, and the atlases map 32% of the trace carriers.