Yes, 2024 is shaping up a big year for robotics!
Introducing
@CovariantAI
's RFM-1, which just like Sora can generate video, but RFM-1 does it for robotic interaction with the world.
But there is so much more it can do. RFM-1 is a multimodal any-to-any sequence model. RFM-1
Thrilled to be teaching a new course on Deep Unsupervised Learning with
@peterxichen
(ImprovedGAN, InfoGAN, PixelCNN++, VLAE, PixelSNAIL, Flow++),
@hojonathanho
(Flow++, GAIL),
@aravind7694
(Flow++).
Follow along here:
Lecture vids go up once captioned
Very excited to share the launch of my new podcast, The Robot Brains. The first episode was just released. My first guest is brilliant AI researcher, educator, and Director of AI and Autopilot Vision at Tesla
@karpathy
!
Want to learn more about GAN, DC GAN, ImprovedGAN, WGAN, WGAN-GP, Progr.GAN, SN-GAN, SAGAN, BigGAN(-Deep), StyleGAN-v1,2, VIB-GAN, GANs as Energy Models?
Vid:
Slides:
Colab:
w/
@Aravind7694
@alexlioralexli
Congatulations
@karpathy
!!!! Teaching the world Deep Learning, Pioneering Image Captioning w/
@drfeifei
, Founding
@OpenAI
, Pioneering AI Self-Driving at
@Tesla
w/
@elonmusk
. What a career already, and you have only just gotten started!!
cc
@techreview
I couldn't be more honored. So thankful for all my amazing collaborators that made this possible, my students and colleagues at
@UCBerkeley
, collaborators at
@OpenAI
and
@CovariantAI
, and of course my MS and PhD advisors
@DaphneKoller
and
@AndrewYNg
, where my AI journey started.
Our warmest congratulations to Pieter Abbeel
@pabbeel
, recipient of the 2021
#ACMPrize
for contributions to robot learning, including learning from demonstrations and deep reinforcement learning for robotic control.
Learn more about Abbeel’s work here:
Excited to share a 6-lecture series on Foundations of Deep RL:
MDP basics, value & policy iteration, max-ent, DQN, policy gradient, TRPO, PPO, DDPG, SAC, model-based RL.
And went in rabbit hole of fixing captions, way too much time, but enjoy the read 😅
Very honored to share that I am the 2022 recipient of the prestigious
@IEEEorg
Kiyo Tomiyasu Award.
A big thank you to all my collaborators, mentors, and nominators/letter writers!
(all 2022 IEEE Technical Field Award winners listed here:
)
One of my favorites from most recent offering of CS287 Advanced Robotics?
Exam study handout summarizing all the main math in ~20pp. Incl. MaxEnt RL, CEM, LQR, Penalty Method, RRTs, Particle Filters, Policy Gradient, TRPO, PPO, Q-learning, DDPG, SAC,
Been looking to learn about CBOW, Glove, Word2Vec, Elmo, GPT, BERT, RoBERTa, ELECTRA, T5, GPT2, and, more generally, learning language models?
Here is a wonderful lecture by
@OpenAI
's
@AlecRad
covering all of these, and more:
What makes production ML hard?
- Cleaning, labeling, and augmenting data
- Troubleshooting training and ensuring reproducibility
- Deploying models and monitoring their real-world impact
To help, we're excited to announce our online production ML course:
Been wanting to learn about CPC, CPC-v2, MoCo, SimCLR, MoCo-v2, and other self-supervised learning techniques?
Wonderful lecture by
@Aravind7694
for the
@UCBerkeley
Deep Unsupervised Learning class:
Exciting announcement: in my
@therobotbrains
conversation with
@geoffreyhinton
(yes!!!!!!!!!), there will be some time for (prepared) audience questions. By Tuesday, post/like questions you have for Geoff in reply to this tweet, and I'll try to incorporate into our conversation.
.
@nvidia
Jensen Huang has always been a big supporter of the research we do at Berkeley AI Research lab and also back when we were at OpenAI. It was really nice to now get to show him at
@CovariantAI
what our Robotics Foundation models are starting to enable in the real world!
My slides from today's Reinforcement Learning in AI event at Stanford:
. Includes a bunch of pointers on getting started towards the end.
Thanks
@AndrewYNg
for organizing, and was a fun panel together with Sergey Levine and
@tydsh
All my
#NIPS2017
slides in one place!
Keynote main conf:
Meta-learning symp:
Robot Learning ws:
Learning in Transportation ws: [w/Greg Kahn]
Hierarchical RL ws:
Last semester, a guest lecture fell through last minute. So improvised a lecture on the backstories behind some research papers -- i.e. not about the research results, but about how we ended up doing research in that direction.
#metaresearch
Online now:
Want to AutoAugment w/o original Google paper's budget to spend?
Population-based Augmentation matches the performance -- with 1000x less compute! :)
BLOG:
PAPER:
w/
@peterxichen
, Daniel Ho, Eric Liang, Ion Stoica, R Liaw
Been looking to learn about self-supervised learning--denoising AE, context encoders*, rotation prediction*, jigsaw, CPC, MoCo, SimCLR*?
Colab with tutorial-demos* now released:
(by
@WilsonYan8
)
Full lecture: (by
@Aravind7694
)
The Full Stack Deep Learning Bootcamp was a lot of fun in person, but of course not everyone can make it in person. Very excited to start releasing the materials today, here:
Happy learning from home!
--
@pabbeel
@josh_tobin_
@sergeykarayev
w/
@l2k
Video lectures of Day 2 of now posted! Includes Troubleshooting, Testing and Deployment, Research Directions, and guest lectures by
@jeremyphoward
and
@RichardSocher
!
Very excited to start offering Full Stack Deep Learning as a formal course at Berkeley this Spring, while also allowing anyone outside of Berkeley to learn along!
If you want to follow along, sign up for updates below!
w/
@sergeykarayev
&
@josh_tobin_
1/ FSDL helps you turn ML experiments into shipped products with real-world impact.
This Spring,
@josh_tobin_
@sergeykarayev
&
@pabbeel
are teaching an improved version as an official Berkeley course:
Want to follow along as we post lectures publicly?👇
ICLR accepted paper analysis by Arthur Pajot. Berkeley coming in second after Google, followed by Stanford, CMU, Facebook, Microsoft, Oxford, IBM, Toronto, ETH, MIT, ... Many other breakdowns, too + code
Gave a lecture on Diffusion Models yesterday with
@wilson1yan
and
@kvfrans
Thanks to it being Lecture 6 into the semester (rather than Lecture 1), we could even include Sora, Mobile Diffusion, SDXL-Lightning :)
Just released Colab accompanying the lecture on Autoregressive Models:
Contains 9 demos/code (see image), including MADE,WaveNet,regular/GrayScale/Parallel PixelCNN
All code by Wilson Yan!
Full course here:
69 = number of invited talks in 2018. Most of these were tailored to specific private, non-technical audiences (so I haven't publicly shared those slides as I have often done for my technical talks), but had some extra time today to clean up:
Very excited to share the news of our 40MM B-round led by
@IndexVentures
w/AI focused
@radicalvcfund
+existing investor
@AmplifyPartners
!
In times when supply chain resilience/robustness couldn't be any more important, excited to take the next steps in the AI Robotics journey!
We’ve raised $40m in Series B funding led by
@IndexVentures
w/ AI-focused
@Radicalvcfund
+ existing investor
@AmplifyPartners
. Grateful for the support of our investors, customers + partners as we continue to bring AI Robotics to the real world!
SO EXCITED to share that
@getonyxfit
has publicly launched! 🎉
Everyone who has spent time in my lab, knows how much importance I put on exercise, and not just for myself :)
Check it out!
Congrats to
@asafagus
@_jamessha
!!
Honored to be small contributor to the Onyx team :)
Excited to share the epic Season 1 finale of
@therobotbrains
with the amazing
@ilyasut
Co-Founder/Chief Scientist
@OpenAI
!
His breakthroughs include AlexNet, seq2seq, MT, GPT, CLIP, DallE,Codex.
"According to my parents, I've been talking about AI at a relatively early age" 😂
Wonderful visit to NYU yesterday! Talk slides:
Also lucky timing I was there same day as
@ylecun
's Turing Award celebration! There was more than one cherry on the Le Cake :)
Deep Q learning requires many tricks to not diverge.
We build on studies of stable (linear) function appr. (Gordon95: , Tsitsiklis,van Roy97: ) through contractions, and extend to NNs + new, stabler "preQN"
Very excited to share my CVPR keynote from this past week:
I share my thoughts on how we might be able to achieve large pre-trained neural networks for robotics, hopefully getting us towards much the way pre-trained models like GPT-x/BERT in NLP.
What an exciting day! Got to personally brief His Majesty, King Philippe of Belgium on Artificial Intelligence. So genuinely kind and thoughtful. And so engaged in our discussion!
Can't share pictures from inside, but beautiful palace, and me in a suit! :)
Been looking to learn more about Semi-Supervised Learning or Unsupervised Distribution Alignment?
One hour tutorials, respectively by
@Aravind7694
and
@peterxichen
, for CS294-158:
Some big news today: we /
@CovariantAI
raised 80MM C-round, led by
@IndexVentures
, and bringing in several new investors!
So much exciting AI Robotics work ahead of us :)
I believe ultimately most of the challenges towards value creation will be in the AI. But setting that aside, what's the hardest R&D ahead for hardware?
I am gonna say "human-level hands". But I think it's feasible, and I for one can't wait to get to work with some of those
Video of Lecture 1 of CS294-158 Deep Unsupervised Learning has been released
We cover:
- Course Logistics
- Motivation
- Autoregressive Models: MADE, Wavenet, PixelCNN(++), pixelsnail (self-attention)
Slides on website
We release ProMP: Proximal Meta-Policy Search. Builds on MAML, E-MAML, DiCE, includes code for all.
Code:
Docs: .
Paper:
with Jonas Rothfuss, Dennis Lee, Ignasi Clavera, Tamim Asfour
Model-based RL, while sample-efficient, often doesn't match performance of model-free RL. Model-based RL via Meta-Policy Optimization resolves this by meta-RL on an ensemble of learned models.
w/ I Clavera, J Ruthfuss, J Schulman, Y Fujita, A Tasfour
Ep.3 of The Robot Brains Podcast is now live! I was fortunate enough to sit down with
@ylecun
. Yann is professor at NYU, Chief AI Scientist at Facebook, winner of the Turing Award (aka Nobel Prize of Computer Science).
Distinct memory from first edition of Deep RL workshop, back in 2015: room so overcrowded the security guard couldn't let Rich Sutton in...
This time, our room (West Exh. Hall C) should have space for everyone! :)
#NeurIPS2019
Super-excited to kick off S3 of
@therobotbrains
podcast with Yoshua Bengio. We discuss LLMs, Higher-Level Cognition, Causality, GFlow Nets, Responsible AI, Human Creativity.
Amazing, very accessible CS Theory talk series I just found out about today. Videos of all lectures available. Just watched Scott Aaronson and Mary Wootters' talks.
Want to learn more about compression and the role of generative models?
Covers: Kraft-McMillan, Shannon, Huffman, Arithmetic Coding, Asymmetric Numeral Systems (ANS), Bits-Back, coding with Autoregressive, VAE, Flow models
Thank you
@geoffreyhinton
for diving deeper into the major potential risks you see with AI, and also reminding us of the tremendous potential for AI to improve our lives.
Video of Lecture 2 of CS294-158 Deep Unsupervised Learning is up
We cover:
- Autoregressive Models: MADE, Wavenet, PixelCNN(++), pixelsnail (self-attention) [ctd]
- Lossless Compression
- Flow Models: NICE, RealNVP, AF, neural AF, IAF, Glow, Flow++
One of our key values at
@CovariantAI
is to always be learning. As part of this I have much enjoyed explaining the core ideas behind recent AI Robotics progress to our customers and partners.
Had some fun producing this with some new animations:
New blogpost! Transformers from scratch.
Modern transformers are super simple, so we can explain them in a really straightforward manner. Includes pytorch code.
Very excited to share our
#ICML2021
tutorial materials on Unsupervised Learning for RL:
w/
@AravSrinivas
Part I covers Representation Learning for RL.
Part II covers Unsupervised Pre-Training for RL.
NIPS Deep RL Workshop is happening again in 2018! Paper submission deadline is Friday Oct 12th. Exciting invited speakers line-up:
@ylecun
, Satinder Singh, Sham Kakade,
@jeffclune
, Doina Precup, Martha White,
@jacobandreas
!
"Our results include using an entire NVIDIA DGX-1 to learn successful strategies in Atari games in single-digit minutes." (thanks
@nvidia
!) Also snuck in a 400K on seaquest (Fig. 4) :p Accelerated Methods for Deep Reinforcement Learning led by Adam Stooke
Biggest pro-tip I got in 2023 (h/t
@RichardSocher
):
Keeping two dumbbells right on my desk, for quick lifts in little dead moments between meetings, when listening to a seminar, etc.
Went from when will I ever have time to lift again to lifting multiple times every day
Couldn’t be more excited about what we are building!
Yet, even more excited about
@CovariantAI
team, all brilliant and exceptional at working across team boundaries, engineering, research, business, we are all having fun and getting it done together!
After 2+ years in stealth, we’re excited to launch today!
Thank you to our team, customers, partners and investors, we couldn’t have done it w/o your support and trust.
Exciting milestone, even more exciting journey ahead!
Current works are restricted to short sequences of texts and images, limiting their ability to model the world.
Presenting Large World Model (LWM): capable of processing long text, images, videos of over 1M tokens (and *no* lost in the middle!)
Project:
We are excited to share Large World Model (LWM), a general-purpose 1M context multimodal autoregressive model. It is trained on a large dataset of diverse long videos and books using RingAttention, and can perform language, image, and video understanding and generation.
Love the content and how this is done! Great insights into how
@berkeley_ai
Prof. Mike Jordan thinks about the future of AI and also a look into his research style
And thank you Barbara Rosario for creating!
Existing RL methods can be adapted to learn robust control policies capable of imitating a broad range of example motion clips, while also learning complex recoveries.
paper: w/
@jasonpeng0
Sergey Levine
@Mvandepanne
On Ep20 of
@therobotbrains
, I sit down with the amazing
@drfeifei
, Co-Director of
@StanfordHAI
, Member of Nat'l Acad. of Medicine
She shares ImageNet origin stories, her interest in physics leading to AI, her work in AI for Medicine, on the Nat'l AI Research Resource Task Force
#nips2017
Meta Learning Symposium videos are now available here:
Thanks to Risto Miikkulainen, Quoc Le, Kenneth Stanley, and Chrisantha Fernando for organizing!
Second episode of The Robot Brains podcast is live now! I was lucky enough to sit down with Princeton Professor
@orussakovsky
and dive into many of the possible issues with the data powering AI systems and what led her to start
@ai4allorg
!
with real-world robotic experiments still more costly and time-consuming than simulated experiments, really exciting to see these major mujoco updates!
Congratulations Dr Rocky Duan!!! Hard to believe how much you got done, and all in just over 2 years!
"Meta Learning for Control"
Thrilled to still be working together, putting AI/robotics into manufacturing and logistics at :)
We (led by
@peterxichen
) introduce a new generative model architecture combining causal convolutions with self attention. New SOTA log-likelihood results on CIFAR-10 (2.85 bits per dim) and 32×32 ImageNet (3.80 bits per dim). and