Read on Twitter
Looking forward to reading this work linking neural language models from NLP to the brain (ECoG). This is an exciting new area (that we've also tried to contribute to) but also one w/ many potential pitfalls.
I'll be asking some constructive/critical questions while reading!
https://twitter.com/HassonLab/status/1334701677253980160
I'll be asking some constructive/critical questions while reading!
https://twitter.com/HassonLab/status/1334701677253980160
(1) Does it report the amount of variance in the neural signal explained by the model? There are many ways to report accuracy (correlation, % of the 'noise ceiling', repres. similarity score, leave-2-out), but those often obscure how small the Explained Variance really is.
(2) That humans and neural language models can accurately predict next words is very well known, but observing when and where the brain predicts the next word is cool. But are the right controls in place to distinguish between brain predicting vs read-off model predicting?
(3) Existing work on relating neural language models to brain imaging is published in very diverse venues. Are they citing the relevant papers? At a first glance NLP/ML venues seem overlooked: no cites to Tom Mitchell [1] and colleagues [2], Jain & Huth [3], not to our work [4].
(4) Brain plots always look very cool, but are not always meaningul. There always is one 'most predictable' location for each model component (even if the prediction is terrible); how do they make sure this really tells us something about the organization of the brain?
(5) Statistics is difficult, given that the same data are often used in multiple modelling studies, & modellers run there models as often as they like. For 'confirmatory' claims, are statistical tests really valid? Or is the work clearly labeled as 'exploratory'? (ours [4] is]).
[1] e.g., https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0112575
[2] https://psycnet.apa.org/record/2017-27842-003
[3] https://scholar.google.com/scholar?cluster=16680485027101599625&hl=nl&as_sdt=2005&sciodt=0,5
[4] https://scholar.google.com/scholar?hl=nl&as_sdt=0%2C5&q=samira+abnar&btnG=
@samiraabnar
[2] https://psycnet.apa.org/record/2017-27842-003
[3] https://scholar.google.com/scholar?cluster=16680485027101599625&hl=nl&as_sdt=2005&sciodt=0,5
[4] https://scholar.google.com/scholar?hl=nl&as_sdt=0%2C5&q=samira+abnar&btnG=
@samiraabnar
And another exciting sounding paper in this area too add to the reading list (and to all such questions about): https://mobile.twitter.com/alex_ander/status/1336402200378138628
