Eugene Vorontsov @VorontsovEugene

Recent PhD graduate from Mila. Montréal, Québec Joined June 2017

Tweets

51
Followers

885
Following

188
Likes

88

Eugene Vorontsov @VorontsovEugene

4 years ago

Some fun quick work for MICCAI DALI: "Label noise in segmentation networks : mitigation must deal with bias" arxiv.org/abs/2107.02189 youtu.be/iLLOwTtAfBc Segmentation models are: - Robust to unbiased errors. - Sensitive to biased errors.

1 1 0 0 1

Download Image

Eugene Vorontsov @VorontsovEugene

4 years ago

Seems my advisor @chrisjpal went from (arcade) street fighting to fighting reviewer #2 in the Canadian AI mafia.

SF_Siberian @Siberian_644

4 years ago

Seems my advisor @chrisjpal went from (arcade) street fighting to fighting reviewer #2 in the Canadian AI mafia.

191 2K 7K 0 462

Download Image

0 0 1 0 0

Eugene Vorontsov @VorontsovEugene

5 years ago

Haven't used this for a while. Thinking maybe I should say I've graduated in the meantime ;)

0 0 4 0 0

Eugene Vorontsov @VorontsovEugene

6 years ago

Check out these big GANs at @BorealisAI

0 0 9 0 0

Download Image

Eugene Vorontsov @VorontsovEugene

6 years ago

#manifencours

1 0 0 0 0

Download Image

Eugene Vorontsov @VorontsovEugene

6 years ago

New: stable but more expressive than an orthogonal RNN, via Schur decomposition. arxiv.org/abs/1905.12080 Pleasure helping with this. Was surprised: independent idea by a new team in the same lab, 3 years after @chiheb_tr and I tried it without all the theory (and didn't pursue).

1 1 14 0 1

Eugene Vorontsov @VorontsovEugene

6 years ago

Not research related but gonna plug my new kitten instagram : instagram.com/cleoandthunder Kittens for when you need a break haha. So many more photos to sort!

0 0 5 0 0

Eugene Vorontsov @VorontsovEugene

6 years ago

We boost segmentation performance for applications such as medical imaging. Image translation for images with <--> without target. Plan to extend to multi-domain (>2): E.g. shadow segmentation. Code planned for release. arxiv.org/abs/1904.01636

1 3 12 0 0

Download Image

Greg Brockman @gdb

6 years ago

I asked @ilyasut how to set neural network init. He accidentally replied with a poem: You want to be on the edge of chaos Too small, and the init will be too stable, with vanishing gradients Too large, and you'll be unstable, due to exploding gradients You want to be on the edge

21 213 1K 0 45

Eugene Vorontsov @VorontsovEugene

7 years ago

We finally published this recently: SOTA auto liver cancer segmentation is poor but manual correction yields near expert-level results, significantly saving time. Thorough eval including intra- and inter-operator variability!

Radiology: Artificial Intelligence @Radiology_AI

7 years ago

1 7 15 0 1

Download Image

1 2 24 0 1

Christopher Beckham, PhD @chris_j_beckham

7 years ago

I don't know if anybody's done this before, but I was curious to see if it was possible to train a 'binary GAN' where the prior p(z) is a Bernoulli(p). The only trick I had to do was scale z := (z-0.5) / 0.5 so that it was zero-centered. Here I'm mapping a 128-dim code to a face

14 119 569 0 42

Download Video

Eugene Vorontsov @VorontsovEugene

7 years ago

We've achieved highest squat density at the Borealis booth.

Christopher Beckham, PhD @chris_j_beckham

7 years ago

We've achieved highest squat density at the Borealis booth.

0 7 14 0 0

Download Image

0 0 3 0 0

Chris Pal @chrisjpal

7 years ago

An opportunity to join Mila as a tenure-track professor in AI at Polytechnique Montréal: polymtl.ca/carriere/en/of…

0 8 10 0 0

Christopher Beckham, PhD @chris_j_beckham

7 years ago

We just released code for our paper 'Unsupervised Depth Estimation, 3D Face Rotation and Replacement', in which we are able to infer depth from faces in a completely unsupervised manner arxiv: arxiv.org/abs/1803.09202 code: github.com/joelmoniz/Dept… video: youtube.com/watch?v=h_brJW…