Memory-Enhanced Sequential Variational Transformer for Inferring Neural Population Dynamics

As deep learning continues to expand its applications in neuroscience, studies on the dynamics of neural populations such as LFADS seek to infer the latent factors from the high-dimensional sequential features. However, these methods neglect the effects of the historical states on the current states of neurons. Therefore, we propose an innovative memory enhanced sequential variational transformer (MESVT), which combines the current states of neural population features with the historical states through the cross-attention mechanism. Furthermore, we validate the performance of MESVT on two collected single-cell-level recording datasets. Utilizing MESVT, we analyze the potential dynamic encoding mechanism of neural populations in the VISP (Ventrolateral Intraparietal Sulcus) for visual stimuli.