🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Amazing work by Aniket and MLO team at GRI. Aniket will present 2 really nice results at COLT and he presented one very strong paper at Neurips last year....1/2

In the age of large-scale models, how does interpretability scale with data and model size? We investigate the effect of scale on mechanistic interpretability by collecting >120’000 human responses and find no improvement! 1/6 🌐brendel-group.github.io/imi 📃arxiv.org/abs/2307.05471

Thrilled to announce Med-PaLM, our medical large language model is published in nature, today. Extremely excited for the possibilities this unlocks! rdcu.be/dgGfJ

Excellent article by Sander Dieleman about diffusion models! My favorite part is about the link to RNNs: “Diffusion models present a way to train deep RNNs without backpropagating through the recurrence at all, yielding a much more scalable training procedure.” sander.ai/2023/07/20/per…

1/Can Neural Network Memorization be Localized?🔍Learn about "Example-tied Dropout" that directs memorization to fixed neurons that can be thrown away at test time! #323@10:30 today #ICML w/Mike Mozer Hanie Sedghi @ZacharyLipton @ZicoKolter Chiyuan Zhang 🌐tinyurl.com/mem-drop

Happy to share our recent #ICML2023 work -- we hope & believe this opens up interesting new directions in supervised learning. Please share!

when your first paper gets demolished but you keep submitting

Nature Comms paper: Subtle adversarial image manipulations influence both human and machine perception! We show that adversarial attacks against computer vision models also transfer (weakly) to humans, even when the attack magnitude is small. nature.com/articles/s4146…

Exciting work by Ramnath Kumar, Dheeraj Nagaraj, Arun Suggala. Just two lines of change leads to significantly more robust learning, and non-trivial gains in a variety of domains including for GLUE benchmarks, Imagenet, OOD datasets, Tabular datasets etc . [1 out of 2]

Super excited about Matformers! It provides ability to train one large model and read-off multiple smaller models without additional training! Slides from our talk at ICCV Workshop on Resource Constrained Deep Vision: prateekjain.org/publications/s… prateekjain.org/publications/s…

Handing over the “big improvements” to your co-author

Legend has it that Americans have 50 words for "breakfast cereal", including one to refer to the transient state of corn flakes where they've soaked up enough milk that they're not too dry, but not yet too soggy, just right

“The establishment of the School of AI at #IITM will help India become a global leader in AI" says Mr. Sunil Wadhwani, an IIT Madras alumni who donated Rs. 110 Cr. to establish a School of Data Science & AI. Watch NDTV youtube.com/watch?v=10Uk9w… J Sam Daniel Stalin Balaraman Ravindran WISH Foundation India

Check out our blog post on early readouts as a means of identifying and mitigating spurious features in neural networks!

Have you ever done a dense grid search over neural network hyperparameters? Like a *really dense* grid search? It looks like this (!!). Blueish colors correspond to hyperparameters for which training converges, redish colors to hyperparameters for which training diverges.

Lumiere is a space-time diffusion research model that generates video from various inputs, including image-to-video. The model generates videos that start with the desired first frame & exhibit intricate coherent motion across the entire video duration → goo.gle/47WX6C2

Excited to share Tandem Transformers: a simple but effective technique to significantly reduce latency of generative LLMs (in some cases ~2.74x). arxiv.org/abs/2402.08644 [1/n]

Check out our blog post on learning under concept drift! We learn reweightings of training data to model information decay and maximize future performance.

Come attend the awesome Bangalore Cognition Workshop at IISc from June 15-21 2024! Apply at forms.gle/kfA1obX8CkPgCP… Deadline: Feb 29 2024 Please circulate widely!

Part of my top secret mission to excite more people to work on enabling Human-Human collaboration through AI !