Signal Studio is in public beta. Usable, but still moving. Join the beta
Docs / Models

Inputs and outputs

What goes into a model card and what comes out.

Inputs

Usually Epochs, sometimes Raw or a Table of features. The manifest declares which, and the type check refuses the wrong one before anything runs.

Shape matters

A model trained on 64 channels at 256 Hz will accept 128 channels at 1024 Hz without complaining and return nonsense. Model cards should declare their expected sampling rate and channel count as parameters, and refuse when they do not match.

If a model card you install does not check this, that is worth reporting to its author. Silent acceptance of the wrong shape is the most common way a model produces a confident wrong answer.

Outputs

TableOne row per trial or per channel, with the prediction and usually a confidence
MatrixChannel by channel, for connectivity or similarity models
EpochsWhen the model transforms rather than classifies

On confidence values

A confidence is a number the model produces, not a probability that it is right. Treat it as a ranking, not as a calibrated measure, unless the author says otherwise and shows the calibration.

Edit this page on GitHub