Fig. 7: Forward transfer of the PFC-MD model with the task similarity measurement.
From: Rapid context inference in a thalamocortical model using recurrent neural networks
A To enable the PFC-MD model with the ability of forward transfer, the MD gating of different tasks is overlapped on some PFC neurons so that cognitive tasks share a subset of PFC neurons. B The task similarity is measured by the inner product of task variance, which is obtained from one recurrent neural network with interleaved training. C The mean performance of the PFC-MD model with different MD-to-PFC effects, and similar and non-similar task pairs to learn. The disjoint PFC-MD model didn’t have forgetting for learned tasks regardless of task similarity but had no forward transfer, while the overlapping PFC-MD model obtained performance improvement in terms of forward transfer. D The mean continual learning (CL) and forward transfer (FT) performance with varying task similarity in the disjoint and overlapping MD-to-PFC projections. The error bars denote the standard deviations.
