Another very interesting paper from the same lab shows how reducing this texture bias can significantly increase robustness and accuracy: https://openreview.net/forum?id=Bygh9j09KX …. It's an oral at ICLR19.
"Approximating CNNs with Bag-of-local-Features models works surprisingly well on ImageNet" https://openreview.net/forum?id=SkfMWhAqYQ … cool/fun paper. A "bag of words" of nets on tiny 17x17 patches suffice to reach AlexNet-level performance on ImageNet. A lot of the information is very local.
-
-
-
Excellent paper! It exposes current CNN architectures limitations, by being biased towards texture instead of shape. Shape biased recognition will improve accuracy/robustness in many areas, including 3D reconstruction from stereo/multi cameras.
- 1 more reply
New conversation -
-
-
Author here. One message of the paper is that we need to be careful: the availability of many weak & local statistical regularities can be sufficient to solve the task in which case DNNs do not learn the underlying "physics" of the world (like object shape). We need better tasks.
- 1 more reply
New conversation -
-
-
Recently watched Arrival movie based on Ted Chiang story. The aliens there effectively used kind of BoW language - they didn't have concept of time nor past/future and didn't organize "words" in "sentence" in any sequential order.
-
Several Indian languages, eg., Sanskrit, Kannada do not have word order. The words arranged in any order give rise to a sentence with the same meaning.
End of conversation
New conversation -
-
-
We worked on something similar that made it to a NIPS poster session, my colleague worked out a lot of the details of what the value of non-relevant feature values mean for feature density in the training set (see appendix) https://arxiv.org/abs/1705.08078v4 …
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
I wrote a more easily digestible summary of the main findings @ https://medium.com/me/stats/post/f4229317261f …
- 1 more reply
New conversation -
-
-
Fascinating... very fascinating. I've kind of vaguely wondered about this before... but if this is the case, I'm super surprised we'd only discover this in 2019. ?!?
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
also on that topic https://openreview.net/pdf?id=Bygh9j09KX …
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
This is really interesting - it reminds me a bit of this paper toohttps://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1006613 …
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
In NLP and applications (IR, IE, classification, MT), Bag-of-Words and Bag-of-NGrams out-did the traditional, symbolic approaches with grammars, syntax, and long-distance dependencies. Interesting to see how similar local features are doing the same in image object recognition.
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
Sounds like it could be a thing
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
Reminds me of some years ago when it was popular to use some kind of regularized SVD on image patches for tasks like inpainting and denoising
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
This makes a lot of sense to me. I've always viewed CNNs from a Bag of Words lens, just with many more features applied that are also learned and not decided upon by a human
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
@RochanAvlur could try this insteadThanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
At first I thought you meant classification of CNN news items :)
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.