A. Description of learning rules corresponding to different types of learning problems and corresponding expressions for the recall factor W ·ΔW used in the recall-gated consolidation model.
B. Schematic indicating a possible implementation of the model in a supervised learning problem, where LTM plasticity is modulated by the consistency between STM predictions and ground-truth labels.
C. Simulation of a binary classification problem, N = 4000, λ = 0.2, θ = 0.125, p = 0.1. There are twenty total stimuli each associated with a random binary (±1) label. Plot shows the classification accuracy over time, given by the outputs of the STM and LTM of the consolidation model. Shaded region indicates standard deviation over 1000 simulations.
D. Like B, but for a reinforcement learning problem. LTM plasticity is gated by both STM action confidence and the presence of reward.
E. Simulation of a reinforcement learning problem, N = 4000, λ = 0.25, θ = 0.75, p = 1.0. There are five total stimuli and two possible actions. Each stimulus has a corresponding action that yields reward. The plot shows average reward per step over time, using the actions given by the STM or LTM.
F. Like B and D, but for an autoassociative unsupervised learning problem. LTM plasticity is gated by familiarity detection in the STM module, learned using a separate set of weights.
G. Simulation of an autoassociative learning problem. N = 4000, λ = 0.1, θ = 0.25, p = 1.0. Recall performance is evaluated by exposing the system to a noisy version of the reliable patterns seen during training, allowing the recurrent dynamics of the network to run for 5 timesteps, and measuring the correlation of the final state of the network with the ground-truth pattern.