1 2 Supplementary Information for 3 Alleviating catastrophic forgetting using context-dependent gating and synaptic stabilization 4 Nicolas Y. Masse, Gregory D. Grant, David J. Freedman 5 David J. Freedman 6 E-mail:
[email protected] 7 This PDF file includes: 8 Supplementary text 9 Figs. S1 to S3 10 Table S1 11 References for SI reference citations Nicolas Y. Masse, Gregory D. Grant, David J. Freedman 1 of9 www.pnas.org/cgi/doi/10.1073/pnas.1803839115 12 Supporting Information Text 13 Network Tasks 14 We trained RNNs on a set of 20 cognition-based tasks, commonly used in neuroscience research. These 20 tasks were 15 implemented in a manner similar to (1). In all 20 tasks, stimuli were represented as coherent motion patterns moving in one of 16 eight, equally spaced, possible directions. However, the results are not meant to be specific to motion, or even visual, inputs. 17 The network output represents a decision to either withhold a response (in many neuroscience experiments, this is equivalent 18 to maintaining fixation), or to respond in one of eight different directions (equivalent to generating a saccade). 19 Stimuli could be presented in one of two locations; neural responses to stimuli in one location were not influenced by stimuli 20 in the other location. A "fixation" cue was also presented, which usually (but not always) indicated to the network to withhold 21 a response. The fixation cue was always presented at the start of each trial, and lasted for a variable amount of time depending 22 on the task contingencies.