Have you ever wondered what are the challenges and opportunities of using neural networks for data prefetching? Title: A Hierarchical Neural Model of Data Prefetching Abstract: asplos-conference.org/abstracts/aspl… @ZhanShi44105240 Akanksha Jain Milad Hashemi Kevin Swersky #ASPLOS21 #NeuralNetwork

Check out new work on ML-driven design and exploration of custom accelerators, showing how #MachineLearning facilitates architecture exploration by rapidly identifying high-performing configurations across a range of applications. Learn more ↓ goo.gle/3rngiEh

Leveraging #MachineLearning for accelerator design enables faster exploration of the architecture search space leading to more efficient hardware across a range of applications. Collaboration w/: Christof Angermüller, Berkin Akin, Yanqi Zhou, Milad Hashemi, Kevin Swersky.

I have a new paper on how to represent part-whole hierarchies in neural networks. arxiv.org/abs/2102.12627

Hi all! Very pleased to share that my latest paper: "Oops I Took A Gradient: Scalable Sampling for Discrete Distributions" (arxiv.org/abs/2102.04509) has been accepted to ICML for a long presentation. Energy-Based Models have seen amazing progress in the last few years...

ICML 2021 Outstanding Paper Award Honorable Mentions: 2/4. Will Grathwohl, Kevin Swersky, Milad Hashemi, David Duvenaud, and Chris Maddison 📜Oops I Took A Gradient: Scalable Sampling for Discrete Distributions (Tuesday 9am US Eastern)

This was a very fun project: an elegant algorithm that works well on the difficult task of sampling from discrete EBMs. Congratulations will grathwohl and team!

Big shout-out to will grathwohl Jackson (Kuan-Chieh) Wang @cvpr jörn jacobsen David Duvenaud Kevin Swersky Mohammad Norouzi for their amazing paper "Your classifier is secretly an energy-based model and you should treat it like one"!

I made a bot improvise for 1000 hours and then asked it to come up with a few short-form improv games of it's own. Here's the first three...

The overall recipe is general, and the same method could be applied to many other design problems in principle! More in the paper: arxiv.org/abs/2110.11346 Awesome collaboration led by Aviral Kumar & Amir Yazdan! w/ Milad Hashemi & Kevin Swersky

📢Introducing Pix2Seq-D, a generalist framework casting panoptic segmentation as a discrete data generation task conditioned on pixels. Works for both images and videos, with minimal task engineering. arxiv.org/abs/2210.06366 work w/ Lala Li, Saurabh Saxena Geoffrey Hinton David Fleet

This was a really interesting project for me to learn about and apply neural fields, with some great collaborators!

This is a really natural framework to improve Bayesian optimization when you have access to related optimization tasks arxiv.org/abs/2109.08215 Joint work with Zi Wang, Ph.D., George E. Dahl, Chansoo Lee, Zachary Nado, Justin Gilmer, Jasper, Zoubin Ghahramani

I’m really excited about this project! Backpropagation and variations are extremely effective at fine-tuning diffusion models on downstream rewards.

Check out Kevin Clark’s and my paper on fine-tuning diffusion models on differentiable rewards! We present DRaFT, which computes gradients through diffusion sampling. DRaFT is efficient & works across many reward functions. With Kevin Swersky, David Fleet arXiv: arxiv.org/abs/2309.17400

We have a student researcher opportunity in our team Google DeepMind in Toronto 🍁 If you’re excited about research on diffusion models, and generative video models, please fill the form : forms.gle/auNq61N35AvTZS… and apply here: deepmind.google/about/careers/…

Greedy Growing Enables High-Resolution Pixel-Based Diffusion Models We address the long-standing problem of how to learn effective pixel-based image diffusion models at scale, introducing a remarkably simple greedy growing method for stable training of large-scale,

🆕🔥We show that LLMs *can* plan if instructed well! 🔥Instructing the model using ICL leads to a significant boost in planning performance, + can be further improved by using long context. arxiv.org/abs/2406.13094 w/ Azade Nova Bernd Bohnet A.Parisi Kati Goshvadi Kevin Swersky Hanjun Dai +

Well, we actually did it. We digitized scent. A fresh summer plum was the first fruit and scent to be fully digitized and reprinted with no human intervention. It smells great. Holy moly, I’m still processing the magnitude of what we’ve done. And yet, it feels like as we cross

Do generative video models learn physical principles from watching videos? Very excited to introduce the Physics-IQ benchmark, a challenging dataset of real-world videos designed to test physical understanding of video models. Webpage: physics-iq.github.io