Fig. 1: Computational modeling approach.
From: Computational models of episodic-like memory in food-caching birds

A In the experiments, jays can eat different items of food, cache them in visuospatially distinct sites (ice cube trays with different arrangements of LEGO Duplo blocks) or inspect their own caches for available food. With different feeding schedules and manipulations of the cached food items the experimenter can change the birds' motivational states and caching experiences. Adapted from Clayton, N., Bussey, T. & Dickinson, A. Can animals recall the past and plan for the future?. Nat Rev Neurosci 4, 685-691 (2003). https://doi.org/10.1038/nrn1180, Springer Nature Limited. B For 28 published experiments we formalized the experimental protocol in a domain-specific language and extracted all measured quantities from text and figures. Our models (top right) describe populations of simulated birds as distributions over dynamical systems. Simulated birds are sampled from these populations to participate in simulated runs of the experimental protocols and produce simulated results. We compute the approximate log-likelihood \(\log \hat{p}\) by comparing simulated with actual results (Methods). The hyperparameters ϑ that determine the mean behavior and the inter-individual differences are adjusted to maximize \(\log \hat{p}\). C In the Plastic Caching Model the preferences of eating ϕeat, caching ϕcache, and inspecting ϕinspect depend on the neural activities and synaptic strengths in a structured neural network featuring motivational control (green), plastic caching weights (red) and an associative memory with systems consolidation (blue). Only some connections are fully drawn; short outgoing lines indicate connections that are suppressed in this figure. The hyperparameters ϑ set initial values of the weights and the speed of learning. D Example experiment 'deKort07 exp4a' (see Figs. S9–S168 for other experiments and all 78 comparison figures per model). Birds that experienced pilfered caches in tray A (blue) stopped caching food in their previously preferred tray A, whereas birds that successfully retrieved their caches in A continued to cache preferentially in A (center, error bars = SEM, n = 4 birds per group). These experimental results are reproduced in simulations with the Plastic Caching Model (left: best out of 105 simulations, error bars = SEM). Also on average (right), the simulated results match qualitatively the experimental data (error bars = average SEM).