Quantifying stimulus-relevant representational drift using cross-modality contrastive learning

The representational drift observed in neural populations raises serious questions about how accurate decoding survives these changes. In primary visual cortex, it is hotly debated whether such variation is a direct tuning shift that would corrupt decoding or if it can be explained by changes in behavioral or internal state, which could be compensated by joint encoding of the stimulus and the state. We estimate the effects of stimulus-relevant representational drift on decoding using a publicly accessible dataset of mouse V1 responses to a natural movie. Because the only invariant component of the sensory experience across all 24 animals is that they all watch the same natural movie, we can learn a subject-invariant efficient neural representation that retains only stimulus-relevant components. We use contrastive learning between the neural response and the stimulus to learn a neural representation for stimulus-relevant features. This learned representation minimizes decoding error as quantified by Bayes risk. We show that it can be used to read out behaviorally relevant stimulus features (time, static scene, optic flow, and joint spatio-temporal features) at 33ms resolution accurately, a finer timescale than what has previously been explored. When we use the model trained on one recording session to derive feature activations on another, decoding performance is reduced by approximately $40\%$. Motion encoding is most susceptible to representational drift. In addition, when we enlarge the error tolerance window to 1sec, we recover stable stimulus encoding across sessions, echoing previous findings. This shows that decoding stimulus features that vary on fast timescales may require complex computation downstream of V1 to compensate for representational drift.