Ivan Skorokhodov
270 posts
Ivan Skorokhodov
@isskoro
- Applying for ML research internships is crazy. Last October, I had 2 MS degrees, 3 first-author papers (iclr/cvpr/iccv) and 4 years of engineering experience, applied in ~15 FAANGs, and the only company I heard back from was DeepMind who rejected me in 1.5 days
- This paper has received significantly less attention than it deserves, so let me shed a bit more light on it and describe why it's so good: 1. It turns out that the classical U-Net image diffusion backbone, which the entire community has been happily building upon during theAnalyzing and Improving the Training Dynamics of Diffusion Models paper page: huggingface.co/papers/2312.02… Diffusion models currently dominate the field of data-driven image synthesis with their unparalleled scaling to large datasets. In this paper, we identify and rectify several
- I've just defended my PhD, and will now get a dog and go live in a forest
- Replying to @isskoroA very bitter "piece of wisdom" which I carried out from this year internships round is that it's meaningless to apply to top-tier labs without a referral: you simply won't be reviewed regardless of your profile
- In the past 1.5 weeks, there appeared 2 papers by 2 different research groups which develop the exactly same (and embarrassingly simple) trick to improve convergence of image/video diffusion models by 20-100+% (sic!) arxiv.org/abs/2502.14831 arxiv.org/abs/2502.09509
- This year, our research team at Snap is hiring multiple Research Engineer interns to help us build large-scale generative models. You don't need to have papers, but should have solid coding and ML skills. It is a good opportunity for BS/MS and early PhD students. Apply at
- Back in the days when I was just starting in deep learning, I got wildly blown away by unsupervised machine translation (starting from arxiv.org/abs/1710.04087). Imagine you travel to an alien planet, sneak into a library, steal all the books, and bring them here on Earth. Now
- The sad truth about modern ML research is that it crazily incentivizes meeting hopping instead of doing actual science. Someone who spends 40 h/week in meetings for 40 different projects will have a more successful career than someone who grinds 40 h/week on a single projectHot take: what if Google Scholar reported two new metrics: (1) median citations per paper and (2) *percent* of papers with 100+ citations? I computed these metrics for some ~200 senior AI researchers: see bit.ly/google-scholar…. The top researchers by median citations per paper
- arxiv.org/pdf/2406.02507 — a new (unusually simple and short) paper by Karras et al. with a kinda crazy idea for a much better diffusion guidance technique. It is similar to CFG, but instead of an unconditional model, they use a smaller version of the base diffusion itself (i.e.,
- Replying to @isskoroWhat do I need to pass screening? Turing Award?
- Our internship applications at Snap for 2024 are now open! snap-research.github.io/cv-call-for-in…. We cover a wide range of topics from multi-modal LLMs and efficient DL to image/video/3D generation and editing
- Replying to @isskoroWhat "saved" me in the end, is that in 2 of these companies there were people who were interested in my research outside of my application flow. They helped me to get through the screening and that led to 2/2 offers
- Today should be my first day as a research intern at Meta. But guess what? I didn't get the visa and the internship got canceled.
