Supporting Text S3 for: Ensembles of spiking neurons with noise support optimal probabilistic inference in a dynamically changing environment

Action-predictive activity in macaque motor cortex is also modulated by the expected value of the action. This was demonstrated in [1]. There, experiments were performed where a visual cue indicated the value of an action in a planar center-out-reaching task. The task was similar to the task of Cisek and Kalaska (2005) in the sense that a visual cue indicated two of eight possible target directions for a subsequent movement. It differed in three respects. First, there was no color cue. Instead, after the visual cue, a go-cue appeared and the movement had to be performed immediately. Second, any possible direction pair could occur instead of the restriction to directions with a 180 • offset in the former task. Finally and most crucially, the border style of the cue in each direction indicated probabilistically the number of juice drops for that action. A " low-value " target (disk with thick black border) had a 60% chance of yielding 1 drop, 30% chance of yielding 2 drops, and a 10% chance of yielding 3 drops. For the " medium-value " target (disk without border), the probabilities were 60% for 2 drops, 20% for 1 drop, and 20% for 3 drops. A " high-value " target (disk with thin border) was worth 3 (60%), 2(30%), or 1 drop (10%) of juice. Experiments were performed with a single target (1-target task) and with two concurrent targets (2-target task). After the go cue, the monkey was free to move to the target of his choice (the authors also considered a task where the monkey was forced to move to a particular target by a corresponding cue, but this task is conceptually similar to the free task with one target). For conceptual simplicity, we redefine the task such that the value (1 to 3 drops of juice) is converted to a probability of receiving a binary reward. The probability of receiving the