Figure 5.
Memory transformation can be optimized to the temporal dynamics of the environment. A, Illustration of the foraging task with randomly sampled bout lengths (the end of a bout is indicated by a jump in the reward location). B, Comparison of the accuracy of episodic versus schematic memories over time. If bouts are sampled from an exponential distribution, then the time it takes for the memory output from the schematic system to more accurately match an analytically computed expected reward distribution is dependent on the rate parameter for bout sampling (βbout, inset). A negative value in the Kullback–Leibler divergence (KLD) difference score indicated better schematic match, whereas positive values indicated better episodic match. C, Test of the ability of a combined network to optimize its performance under conditions where bouts were sampled using different βbout values. Larger circles represent the best performing episodic-schematic switch value (βα), with other values being compared with this best performing value. As the bout lengths increased (blue vs green vs yellow circles), the optimal episodic-schematic switch time similarly increased.