Neuroquery: Input terms in neuroquery don't have any weights in generated brain map?

1

I want to use Neuroquery to get a statistical image for the term “central executive network”. Neuroquery recognizes the terms “central executive” and “network”, but when I look at the weights, neither of these tokens has any influence on the brain map? Instead, Neuroquery seems to generate the brain map purely from related terms like “dmn”? Or am I perhaps misinterpreting the diagrams?

Maybe @jeromedockes can help here?

image
image1269×736 59.3 KB

2

Yes, your understanding is correct. In that case, it means that the image you’ll get is probably not a reliable image of the CEN.
Best,
Bertrand

1 Like
3

Hi Betrand, do you know why this is the case? Looking at the user interface, I can see that Neuroquery seems to find 353 studies with the term “central executive”. Then why are these studies not included to predict the brain map? Of course, if I were to type in “foobar”, that makes more sense since the term itself isn’t even recognized as a valid token.

4

The neuroquery estimator has a built-in thresholding procedure. It may be that “central executive” does not survive it. See eq. (12) and following lines in

@jeromedockes can you confirm ?
Best,
Bertrand

5

yes the amount of shrinking applied to different term’s coefficients does not only depend on the number of documents they appear in but on how well they predict the activations in those studies. Here it seems the model did not find a very strong, reliable association between this term and the brain activations so due to the smoothing and thresholding procedure @bthirion mentioned it ends up relying on related terms such as dmn rather than on “central executive network” itself. Basically what it means is neuroquery does not have a good answer for that query. when that happens we try to detect it and show a warning in the interface but the heuristic we use to decide if the warning should appear is not perfect

1 Like
6

Got it! Thanks! Last thought: Wouldn’t it make sense if the algorithm then checks one more time and throws an error (or a warning) for cases like this? When the weights of the input tokens are 0 and the map is purely based on only similar terms?