Abstract
Sensory stimuli arrive in a continuous stream. By learning statistical regularities in the sequence of stimuli, the brain can predict future stimuli (Xu et al. 2012; Gavornik and Bear 2014; Maniscalco et al. 2018; J. Fiser and Aslin 2002). Such learning requires associating immediate sensory information with the memory of recently encountered stimuli (Ostojic and Fusi 2013; Kiyonaga et al. 2017). However, new sensory information can also interfere with short-term memories (Parthasarathy et al. 2017). How the brain prevents such interference is unknown. Here, we show that sensory representations rotate in neural space over time, to form an independent memory representation, thus reducing interference with future sensory inputs. We used an implicit learning paradigm in mice to study how statistical regularities in a sequence of stimuli are learned and represented in primary auditory cortex. Mice experienced both common sequences of stimuli (e.g. ABCD) and uncommon sequences (e.g. XYCD). Over four days of learning, the neural population representation of commonly associated stimuli (e.g. A and C) converged. This facilitated the prediction of upcoming stimuli, but also led unexpected sensory inputs to overwrite the sensory representation of previous stimuli (postdiction). Surprisingly, we found the memory of previous stimuli persisted in a second, orthogonal dimension. Unsupervised clustering of functional cell types revealed that the emergence of this second memory dimension is supported by two separate types of neurons; a ‘stable’ population that maintained its selectivity throughout the sequence and a ‘switching’ population that dynamically inverted its selectivity. This combination of sustained and dynamic representations produces a rotation of the encoding dimension in the neural population. This rotational dynamic may be a general principle, by which the cortex protects memories of prior events from interference by incoming stimuli.