Contrastive neural representation (CEBRA)

A representation-learning approach that shapes an embedding directly by an InfoNCE-style contrastive loss: pull together neural samples that share an auxiliary variable — elapsed time, or a behavioral label — and push apart the rest, without ever reconstructing the activity. CEBRA (Schneider, Lee and Mathis, Nature 2023) is the canonical instance; it can run time-only (fully self-supervised) or hypothesis-guided by behavior, produces embeddings that are consistent across subjects, sessions and even modalities, and comes with goodness-of-fit and consistency statistics rather than a single opaque accuracy number.

Contrastive learning sits opposite the generative and masked family: instead of modeling a likelihood over activity, it optimizes the geometry of the latent space so that distances carry meaning. That makes it well suited to consistency and alignment questions, and it is often used as an embedding front end whose output feeds a simple downstream decoder.