Population-level embedding

A low-dimensional representation of the joint activity of many neurons, learned so that its geometry reflects behavioral or latent variables and is consistent across recordings. In the foundation-model view the embedding is the transferable interface: pretrain it once, freeze it, and attach a light task-specific readout, rather than training a bespoke decoder per session. Methods include contrastive schemes (CEBRA), variational autoencoders (pi-VAE), and the latent bottleneck of a pretrained transformer.

The concept overlaps with the neural-manifold and latent-dynamics literature, but the emphasis here differs. Latent-dynamics work asks what dynamical law governs the state; representation learning asks for an embedding that is transferable, identifiable, and consistent — one whose geometry can be trusted to line up across subjects and days. Identifiability guarantees (up to known transforms) are prized because they let you compare embeddings between animals rather than treating each as an isolated coordinate system.