Quantifying and Modelling Transfer Learning in Mice Between Consecutive Training Stages of a Change Detection Task

Animals are known to be able to rapidly transfer knowledge between tasks with similar structure. We trained a set of mice on a visual change detection task with multiple stages, starting with direct transitions between gratings, adding an intervening gray screen, and subsequently moving to multiple sets of natural images. We observe that, when progressing to new stages, the performance increases very fast. However, when transitioning to a task of higher complexity, the peak performance decreases. This setup facilitates for the first time an experimental platform to study the transfer learning phenomena in mice using visual stimuli. Based on these results and additional neuroscience insight, we propose a cognitive model to explain the quick adaptation observed in mice. It extends a deep Q learning agent with a multi-tiered architecture and the possibility of performing a representation remapping at every level of the hierarchy to prevent the downstream propagation of representation anomalies. This architecture provides the substrate of an adaptation algorithm based on ideas of optimal transport of probability distributions. It matches well key behavioral aspects observed in mice and the experimental constraint that the mice are initially trained using a single task variant before it transitions to the new training phase. The modelling process helped us to gain biological insights: first, the optimal transport mechanism of the representation remapping indicated that a possible reason for the reduced performance when mice move to a new more complex task could be due to limited representation resources which were optimized for the previous task. Second, the multi-tiered architecture was first an engineering constraint and later a biologic insight confirmed by the literature. A final insight came from the computations required to perform the representation remapping. These computations are interesting because they could help to confirm this transfer learning theory by looking for similar neural correlates during the adaptation process.