@michael_diskin profile picture
Michael Diskin @michael_diskin
Joined September 2013
  • Tweets 68
  • Following 272
  • Followers 126
9 Photos and videos
Michael Diskin retweeted
Yandex Research@YandexResearch
14 Mar 2023
Heterophilous graphs, in which edges connect dissimilar nodes, are challenging for GNNs. In our #iclr2023 paper, we propose a new benchmark of 5 structurally diverse heterophilous graph datasets. Read more in this post by Oleg Platonov and @LProkhorenkova research.yandex.com/blog/int…

Introducing new heterophilous graph datasets

We have proposed a benchmark of five diverse heterophilous graphs that come from different domains and exhibit a variety of structural properties. Our benchmark includes a word dependency graph...

research.yandex.com
3
10
1,235
Michael Diskin retweeted
Max Ryabinin
@m_ryabinin
13 Feb 2023
We present SWARM, an efficient algorithm for model-parallel training across the Internet (e.g. with volunteers). Key advantages: 💎 Fault-tolerant ⚖️ Self-balancing on slow GPUs/networks 🐌 Works in low-bandwidth setups 📜 arxiv.org/abs/2301.11913 🖥️ github.com/yandex-research/s…
11
61
263
53,675
Michael Diskin retweeted
Andrey Malinin@AndreyMalinin
10 Dec 2021
Want to assess uncertainty and robustness in Self-Driving, Translation and Weather Prediction? Check out our #NeurIPS2021 paper: Shifts: A Dataset of Real Distributional Shift Across Multiple Large-Scale Tasks: tinyurl.com/2p8833dp See poster B1 at DBT, session 4 room 2
1
6
12
Michael Diskin retweeted
Yandex Research@YandexResearch
1 Nov 2021
At #NeurIPS2021, we will showcase our latest research in a variety of ML-related areas, with a total of eight papers, two benchmarks and one demonstration accepted at the conference. You can find the full details in the link below: research.yandex.com/news/pap…
4
5
I'll continue bragging about our work. We were able to make distributed learning secure without losing the efficiency of the approach from the previous article. A lot of math is included.😉
Alexander Borzunov@sasha_borzunov
3 Aug 2021
Secure Distributed Training at Scale: our latest work proposes a protocol that protects decentralized training of large neural nets from malicious participants with marginal communication overhead. (1/8) Paper: arxiv.org/abs/2106.11257
2
It's hard to brag about your own work. But we have proved here that it is possible to train neural nets distributed, on heterogeneous hardware in unstable conditions. In particular, using volunteers, and Google Colab. And it can even be useful to people. arxiv.org/abs/2106.10207

Distributed Deep Learning in Open Collaborations

Modern deep learning applications require increasingly more compute to train state-of-the-art models. To address this demand, large corporations and institutions use dedicated High-Performance...

arxiv.org
Hugging Face
@huggingface
15 Jul 2021
🧐What if we could use existing computing resources from different entities to train large models? We have shown that is possible by collaboratively training a language model with 40 volunteers🔥🔥 Read our new blog post on our 🤗 website to know more! 🔍 hf.co/blog/collaborative-tra…
0:28
1
2
17
And yes, we are trying to make such distributed training accessible to everyone. Just try it github.com/learning-at-home/…

1
6
Special thanks to @huggingface for their invaluable collaboration!
4
Michael Diskin retweeted
Max Ryabinin
@m_ryabinin
21 Jun 2021
In our latest work, we propose DeDLOC — a method for efficient collaborative training. This approach allowed us to pretrain sahajBERT (a Bengali-language ALBERT) together with the help of volunteers from the community! (1/10) arxiv.org/abs/2106.10207 github.com/yandex-research/D…
3
39
133
Load more