Comparability and Generalization of POCO Unit Embeddings

Determine whether the unit embeddings learned by POCO, a population-conditioned neural activity forecaster trained on calcium imaging data, are comparable across different species and whether these embeddings generalize to brain regions not included during training.

Background

POCO learns unit embeddings that capture meaningful structure in neural data without anatomical supervision, including clustering by brain region within individual datasets. This suggests that the embeddings encode functional properties of neurons relevant for forecasting.

However, despite within-dataset structure, the authors explicitly note uncertainty about whether these embeddings align across species or transfer to brain regions that were not observed during training. Establishing such comparability and generalization would indicate broader utility of POCO’s learned representations beyond specific datasets.