Why are brains the way they are? Are their circuit architectures and synaptic plasticity rules in some sense ‘optimal’? If so, in what sense, or in what contexts? We address these questions using olfactory associative memory in the fruit fly Drosophila. Flies can learn to associate a particular odour with a reward (e.g., food) or punishment (e.g., shock) and thereafter approach or avoid the trained odour. These associative memories are stored in Kenyon cells in the mushroom body, by weakening synapses from odour-responsive Kenyon cells onto mushroom body output neurons (MBONs) that lead to incorrect actions (e.g., odour+punishment weakens KC->Approach synapses). Why weaken incorrect actions rather than strengthening correct actions? Notably, synaptic depression is also used for learning in the vertebrate cerebellum, which has a remarkably similar architecture to the insect mushroom body, suggesting that using depression may be functionally advantageous.

We show both analytically and using simulations that depression outperforms potentiation for discriminating odours with overlapping KC representations, under a particular condition: if behaviour depends on the relative, not the absolute, difference between Avoid vs. Approach MBON activities (i.e. divisive rather than subtractive normalisation). To test whether behaviour depends on the relative difference, we measured aversive learning for a range of odour concentrations and punishment intensities, in an individual-fly T-maze (n=95-442 flies per condition). We automatically tracked the flies’ decisions to enter or leave the side with the punished odour, and from the statistical distributions of these stochastic decisions, we inferred the mean and variance of the flies’ underlying preference for/against the odour. We fitted these data to alternative mechanistic models for learned decision-making that used subtractive or divisive normalisation, and the fit was better with divisive normalisation. These results suggest that flies learn by synaptic depression because, in the mushroom body, it is computationally superior to synaptic potentiation. These results illustrate how quantitative analysis of natural behaviour illuminates neural mechanisms underlying learned decision-making.

N.A., K.G.-W., J.S.J. and M.W.T. contributed equally to this work.