Ankit Dhall

Building generative models using libraries is great for getting started but the high-level APIs abstract away the details (usually where the devil is) I implemented Torchsmith, an open-source package that implements generative models from scratch using basic PyTorch operations🧵

@ottogin1 @CVPR Thanks @ottogin1 for the effort to make it accessible to a wider audience

Hi, Ankit, agree, hah! We are figuring out the logistics of recording with @CVPR, even if for some reason we won't be able to record on the conference, we will make sure to reenact it and upload the videos. In the meantime, check out our MIT mini course where we cover some parts of it: practical-diffusion.org

Join us at @CVPR in Denver for a full-day tutorial about Analytic Understanding of Diffusion Models. The training objective of diffusion models has a closed-form solution -- yet it only memorizes. How do real models generalize? We'll unpack this paradox and the emerging analytical theory behind it. @yuancy @CScarvelis @MasonKamb @WangBinxu @vincesitzmann @JustinMSolomon @SuryaGanguli

I'm excited to start a new journey at @NVIDIA as Senior Deep Learning Engineer working on Visual Generative AI in the Deep Learning Algorithms team!

@bose_joey Will there be a recording?

Heads up, NeurIPS attendees! 👋 🚨Don't miss the FPI Workshop (Frontiers of Probabilistic Inference) back again on Dec 7th. 🔉We've curated an incredible set of accepted papers showcasing the future of probabilistic modeling. Schedule & papers: fpineurips.framer.website We also have a stellar lineup of invited speakers that should be of broad interest to people in AI4Science #neurips2025 #ProbabilisticML #AI Rumours are that there's also a workshop party🫢. co-organized with a stellar team @tara_aksa @msalbergo @marylougab @theh2o64 @guanhorng_liu @k_neklyudov Grant Rotskoff @EvaSmorodina @AlexanderTong7

Scrolling the AI news timeline as a researcher feels like a teenager browsing Instagram: "Everyone else has figured everything out!" Reliable home robots imminent, 100× productivity AI agents, insane visual generation ... Exciting, but anxiety-inducing. What am I doing? 😬

Finally, a single unified text I have been waiting for.

@jm_alexia Was going to request for a recording but you have already taken care of it🙏

I will post the recordings after, don't worry.

I will give a presentation on "Tiny Recursion Models" tomorrow at 1pm in the Mila Agora (6650 Saint-Urbain, Montreal). Its open to everyone, feel free to come by!

Looking forward to reading this one

(5/5) This thread is a concise version, check out the repo on GitHub for the full details github.com/ankitdhall/tor… P.S.: Definitely recommend this way for anyone who wants to get a deeper understanding and see how to train and sample from such models

(4/5) Flow matching: - ODE formulation: Gaussian flow matching and Linear flow matching - SDE formulation: Langevin flow matching

Building generative models using libraries is great for getting started but the high-level APIs abstract away the details (usually where the devil is) I implemented Torchsmith, an open-source package that implements generative models from scratch using basic PyTorch operations🧵

Matrix multiplication is now a multi-trillion dollar business.

it’s been an interesting ride watching the conventional nomenclature of machine learning gradually lose all meaning. there used to be TRAIN and TEST and everything was simple. now we train on the universe. and we test on the universe, too. are we gaming our benchmarks? are we extrapolating or interpolating? if a model is trained on the entire internet but generates a single novel sentence, is it just combining phrases from Reddit or writing something truly novel? feels like we lack the words to even describe the systems we’ve built.

AdamW the Undisputed

@apsarathchandar @3blue1brown +1 for recommending Gilbert Strang's lectures

I am teaching intro to ML this Fall (starting today!), and watching @3blue1brown linear algebra videos several times is a mandatory prerequisite for the students!

while doing ML research i spend a disproportionately large amount of time looking at this specific figure

It's ASML's world and we just live in it

@giffmana @XiaohuaZhai @__kolesnikov__ @_basilM Okay, so the values were obtained via hyper parameter search. Not clear what you mean when referring to "prior". > "This is because the bias term ensures that the training starts close to the prior"

@giffmana @XiaohuaZhai @__kolesnikov__ @_basilM "We initialize t' and b to log 10 and −10 respectively. This makes sure the training starts roughly close to the prior ..." Couldn't find more about how the two values were determined in the text.

What makes CLIP work? The contrast with negatives via softmax? The more negatives, the better -> large batch-size? We'll answer "no" to both in our ICCV oral🤓 By introducing SigLIP, a simpler CLIP that also works better and is more scalable, we can study the extremes. Hop in🧶