Sequential learning in recurrent neural networks create memory traces of learned tasks

Animals can learn and perform many behaviors without interference. However, it is unclear how changes in neural activity during learning avoid interference with activity underlying similar behaviors. Recently, Losey et al. (bioRxiv, 2022) identified a possible explanation by examining monkey motor cortical population activity under a brain-computer interface (BCI) paradigm where linear mappings controlled a computer cursor. The activity produced for a known mapping was altered after learning a new mapping in a way that made the activity more beneficial for this new mapping. That is, the activity retained a “memory trace” of the new mapping without compromising the performance for the already known mapping. Since these memory traces suggest a potential substrate for continual learning, we sought to understand how and when they arise.

To probe the effects of different learning processes on the memory trace, we modeled motor cortical neural activity during the same BCI paradigm using a recurrent neural network. We replicated several experimental results by sequentially training the network on two mappings that required activity in the same low-dimensional subspace, or manifold, representing “within-manifold” perturbations. Intriguingly, we observed memory traces even when network activity was not specifically constrained to maintain both maps. Thus, memory traces may inherently arise from sequential learning. While the presence of these traces was consistent across different BCI maps, their magnitudes differed greatly. Initial changes in behavior imposed by a given mapping were a good predictor of the memory trace’s magnitude, but initial changes in network activity were not. This allowed us to characterize, for the first time, differences in learning between different BCI mappings that represent easily-learnable within-manifold perturbations. Overall, we provide a taxonomy of the neural and behavioral factors shaping continual learning, allowing us to examine potential underlying mechanisms and behavioral constraints in future work.