Fig. 4: How Hebbian plasticity supports context inference in MD.
From: Rapid context inference in a thalamocortical model using recurrent neural networks
A The PFC-MD model was trained in alternating blocks with three contexts. Each context consist of two cues. Three task variables of consecutive trials in block 1 and block 2 are shown: B the MD presynaptic traces, C the MD outputs and D the weight changes over time. The first column is block 1 (Context 1) and the second column is block 2 (Context 2). Suppose that there were three types of MD neurons: MDA, MDB, and MDC. E In the first block, the connections between Context 1 PFC neurons and MDA neurons were enhanced while the connections between Context 2 PFC neurons and both MDB and MDC were enhanced. When Context 2 was presented, MDB and MDC neurons were separated to encode different contexts. The dashed arrow denotes the weakened connection between neurons. F The changes of the PFC-MD synaptic weights over time during model training: synaptic weights PFC  → MD (top) and additive synaptic weights MD  → PFC (bottom). The temporal contexts were encoded sequentially in the MD as incoming contexts were presented in blocks.
