Work in progress — pages and figures are still being written.

Alicia Zeng

All research

Bayesian

Every model here is fit to predict something measured. This area is about what you may then conclude from its parameters.

A model can predict well and still support no conclusion. When its features are correlated, many different settings of the weights produce the same prediction, so the weights a fit returns are one choice among many that the data cannot separate. Accuracy does not reveal this, because accuracy is identical for all of them.

This is a property of the basis, not of the fit. If the features are sparse and non-negative, with one interpretable direction per coordinate, the weights become identifiable, and reading them off is a measurement rather than an interpretation of an arbitrary point.

Sparse concept atoms–

Word embeddings predict cortical responses well. They also carry more semantic factors than they have dimensions, so the factors share directions, and a voxel’s weights cannot be attributed to any one of them. Sparse dictionary learning rewrites the embedding as a thousand non-negative concept atoms, one interpretable direction per atom. Prediction accuracy is unchanged; the difference is that the weights can now be read an atom at a time. The video shows what the models are fit to: measured responses while a story plays.

The talkNeurIPS , five minutes

Five minutes on the same argument: why dense embeddings leave voxel weights unreadable, what sparse dictionary learning does about it, and what the resulting maps show. Recorded at the conference, and also on SlidesLive.

[video]

[New project — name]

[Two or three sentences on what it is and why it belongs here, when it is ready to be described.]