You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Dear Dick Cammarata, Gabriel Goh, Shan Carter, Ludwig Schubert, Michael Petrov and Chris Olah,
I read with interest your paper on curve detectors on distill, lovely work.
I wanted to make you aware that there is a significant amount of effort in Cognitive Science (a branch of psychology that invented neural networks) about single unit 'detectors' and analysing the selectivity and function of these devices.
We have a recent paper on 'object detectors' in deep CNNs which will shortly becoming out in Vision Research which you may be interested in. It's currently up on arXiv.
'Are there any 'object detectors' in the hidden layers of CNNs trained to identify objects or scenes?' Vision Research 2020 https://arxiv.org/abs/2007.01062
The authors, and the Bowers group at the University of Bristol, have been undertaking the study of interpretability, single units and selectivity for years and there are several papers by us and others in this area. There are many references to the missing literature from cognitive science and neuroscience within the lit reviews of these papers.
When you stated 'At the same time, there's a significant amount of skepticism, only partially reflected in the literature.' I think a survey of relevant papers considering this question from cognitive science, psychology and some neuroscience might help expand your argument.
The issue:
I believe that a more thorough literature review including the results from cognitive science on this question, and the single unit 'concept cells' found in neuroscience (discussed at length by Jeff Bowers in his papers) would improve your paper. And the specific paper I mentioned is part of that relevant literature review.
This comment is intended as peer review.
Dr. E. Gale
University of Bristol
The text was updated successfully, but these errors were encountered:
Dear Dick Cammarata, Gabriel Goh, Shan Carter, Ludwig Schubert, Michael Petrov and Chris Olah,
I read with interest your paper on curve detectors on distill, lovely work.
I wanted to make you aware that there is a significant amount of effort in Cognitive Science (a branch of psychology that invented neural networks) about single unit 'detectors' and analysing the selectivity and function of these devices.
We have a recent paper on 'object detectors' in deep CNNs which will shortly becoming out in Vision Research which you may be interested in. It's currently up on arXiv.
'Are there any 'object detectors' in the hidden layers of CNNs trained to identify objects or scenes?' Vision Research 2020
https://arxiv.org/abs/2007.01062
The authors, and the Bowers group at the University of Bristol, have been undertaking the study of interpretability, single units and selectivity for years and there are several papers by us and others in this area. There are many references to the missing literature from cognitive science and neuroscience within the lit reviews of these papers.
When you stated 'At the same time, there's a significant amount of skepticism, only partially reflected in the literature.' I think a survey of relevant papers considering this question from cognitive science, psychology and some neuroscience might help expand your argument.
The issue:
This comment is intended as peer review.
Dr. E. Gale
University of Bristol
The text was updated successfully, but these errors were encountered: