Calcium imaging records continuous fluorescence traces. Many sequence models first discretize these traces into tokens and then train a classifier over a learned vocabulary. This can be effective, but it introduces a discrete interface between the original signal and the temporal model.
The direct alternative
CAPT uses continuous temporal patches as both inputs and prediction targets. The backbone predicts the next patch directly with a regression objective, avoiding a separate vector-quantization stage.
This design is intentionally simple. Once the discrete tokenizer is removed, the natural target is the next continuous segment of the signal. The central empirical question is whether this end-to-end formulation learns representations that remain useful beyond the pretraining dataset.
What matters in evaluation
Forecasting performance is one test, but transfer requires a stricter protocol. In our setting, the pretrained backbone is frozen on target datasets; only neuron/session embeddings or task-specific heads are updated. This separates representation reuse from the benefits of fully retraining a large model.
Open questions
- How far can continuous autoregressive models roll out before correlations decay?
- Which aspects of cross-species transfer reflect common dynamics rather than recording statistics?
- How should pretraining mixtures balance scale, diversity, and biological comparability?
This note is part of an evolving research notebook. Publication claims should be cited from the linked papers rather than from this informal summary.