ts3#

Task Suite 3: Brain Region Prediction.

10-class Cosmos-level region classification (Allen CCF) from individual unit activity, spanning coarse but functionally meaningful regions such as Isocortex, Hippocampus, and Cerebellum. Evaluated in single-unit and multi-unit settings, zero-shot on held-out animals. Two regimes probe different generalization axes: transductive zero-shot (adaptation via non-region supervision) and inductive zero-shot (no adaptation, testing invariance across sessions, probes, and animals). Performance is measured with macro-averaged F1.

Which regime a model can be evaluated under is a property of the model, not a setting: see ts3.models.inductive and ts3.models.transductive.

Dataset#

IBLBrainWideBenchTS3

TS3's view of the benchmark: whole sessions after neural QC.

Extraction#

Extractor

A unit-embedding producer, driven by ts3/extract.py.

ISI histograms#

compute_isi_histogram

ISI histogram, L1-normalized to sum=1.