Latent dimensionality in real neural spiking data

Determine the latent dimensionality of neural population activity in real recordings, i.e., the number of latent variables required to capture shared dynamics in spiking datasets, to enable selection of an appropriate latent dimension when training low-dimensional latent variable models for neural spiking data such as the structured state-space autoencoders used in LDNS.

Background

Low-dimensional latent variable models are widely used to capture shared population dynamics in neural spiking data, but their performance and interpretability depend critically on the chosen latent dimensionality. The authors note that in real neural recordings the true latent dimensionality is not known, making model selection difficult.

Within LDNS, the autoencoder compresses high-dimensional neural activity into a low-dimensional latent space (e.g., 182-to-16 dimensions in monkey data). Although LDNS performs well under such compression, identifying the appropriate latent dimensionality remains challenging and impacts both expressiveness and interpretability.