Representational drift leads to sparse activity solutions that are robust to noise and learning

Memories are thought to be stored in synapses between neurons, but these connections in the hippocampus show complete turnover within a fortnight[1]. Dynamics have also been seen at the representational level, as the cells responding to a particular position change across days even when the animal isn’t overtly learning or forgetting[2]. While some researchers consider representational dynamics to be an inevitable consequence of ongoing learning, others suggest that drift may be beneficial. For instance, theoretical work by Kappel et al.,[3] suggests that if initial learning leads the system to a locally optimal performance solution then drift would enable the brain to not get stuck there and instead find globally optimal solution. This requires drop in performance when the system is exploring low-performance regions between the optima. If there is incomplete exploration of solution space and representations explored by drift are equivalent in performance, then it isn’t known if the solutions explored by drift are in any way better than initially learned solution. Here we modeled drift as uniform sampling of weight solution space and showed that drift leads to sparse activity solutions, which are robust to noise and suitable for continual learning. A threshold-linear activation function produces more weights corresponding to a sparse activity solution than to a dense activity solution[4]. Thus, in simulations drift quickly led to sparse representations with few active neurons for each position in the environment. These sparse activity solutions found by drift expanded the solution space of weights, so they were more robust to noise and could easily learn new information when compared to equally well performing dense activity solutions. Our results demonstrate that sparsity can be a natural consequence of drift, and drift can be used in biological and artificial agents to help learn new memories while retaining old ones.