Ten ways to fool the masses with machine learning

@fayyazhere, @aAmna_01, Asa Ben-Hur

tl;dr:
1) Use an uninformative or irrelevant accuracy metric
2)Inappropriate model selection strategy

Pity that I haven't seen that before doing my talk
1/4
arxiv.org/abs/1901.01686 3)Ignore the fact that examples may not be independent of each other
4) Do not compare with simple baseline classifier
5)Compare your model with un-optimized versions of other models or ones that have
been trained using different data
2/4

Today my wife showed me computer vision app, which solves realworld actual problems: which lego bricks one have and what we can assemble with it .
The recognition is far from ideal, but I appreciate how hard the problem is.
brickit.app
P.S. not an ad :) Moving from MNIST to CIFAR-10

After 16 hours working on a video for presentation, it crossed my mind that it would be a great idea before going to bed, to upgrade my desktop from Ubuntu 16.04 to 20.04 (yes, at 2:30 am).
What could possibly go wrong? To summarize:
1st problem were some binaries from zoom installed.
Then it clashed with libgl -mesa-something. Nvidia drivers got involved

What helped, is running apt -o Dpkg::Options::="--force-overwrite"
2/

This is a nice counter example against my arguments for arXiv.

https://twitter.com/atito_sara/status/1459878452568461314?t=iyuPpTRrPDy-P4pD07YXIQ&s=19

The situation is following, in thread:
1) relatively small group gets the ideas of AE+ViT=SiT, validates on CIFAR, STL,puts on arXiv in April. Also submits to PAMI
1/ 2) AK posts about SiT

https://mobile.twitter.com/ak92501/status/1380322070073118722

3) Masked AE + ViT from Meta AI, comes out in November, gets nice ImageNet results and get all attention. Not citing SiT

So arXiv does not protect SiT from scooping.
Oh, and PAMI rejects SiT
2/

Thanks to @david_picard , I can now formulate why I am for social media.
That is an instrument, where many things depends on you and how good you have done your job.
Writing a good paper is another such thing, but then you are dependent on (random) reviewer choice.
1/ ArXiv is better than the conference in a sense that everyone can see and judge. But it also quite random: your target audience may not check the feed today. However, it is googlable, so not full failure
2/

Let me introduce pixelstitch: simple correspondence annotator based on @matplotlib + jupyter notebook.
Just propvide list of image_fnames and
import CorrespondenceAnnotator.
You can add and erase points, zoom, move, and visualize epipolar geometry, induced by correspondences. 1/ You can install it by
pip install pixelstitch
and the documentation is here:
ducha-aiki.github.io/pixelstitch/
Ofc, powered by nbdev.
All suggestions are welcomed
2/3

I know, you all have dreamed about the conversion functions between #kornia @kornia_foss and OpenCV. Here they are.

github.com/ducha-aiki/kor…

pip install kornia_moons

Docs are here: ducha-aiki.github.io/kornia-moons/

Powered by @fastdotai and @HamelHusain nbdev
1/3 For example, you can detect ORB features is OpenCV and convert them to kornia lafs with

lafs = laf_from_opencv_ORB_kpts(kps)

Or you can match descriptors on GPU with kornia.feature.match_snn() and transfer to OpenCV with cv2_matches_from_kornia 2/3

While I am selecting nice pictures for my PhD thesis, let's do a fun visual benchmark of local features on the photos I took myself.

SIFT-HardNet, R2D2, SuperPoint, SuperGlue
1/ Nothing works here, ha :)

SIFT-HardNet, R2D2, SuperPoint, SuperGlue
2/