Many local minima in Deep Networks are connected through low-loss valleys. Curious about the theoretical origin of such valleys? Check out our
#ICLR2023
Paper on
#symmetry
, local minima and conserved quantities.
Paper:
(1/3)
I have some exciting news to share. After spending some wonderful years at Northeastern, I will be joining
@ucsd_cse
at
@UCSanDiego
as an Assistant Professor in July 2020.
I want to thank all my mentors for their support, and for folks at UCSD for making this happen.
Want to design
#DeepLearning
models that can generalize across dynamics? Check out our
#NeurIPS2022
paper on ``Meta-Learning Dynamics Forecasting''.
Paper:
Code:
(1/3)
How to design Diffusion Models for spatiotemporal forecasting? Check out our
#NeurIPS2023
Paper on ``DYffusion: A Dynamics-informed Diffusion Model for Spatiotemporal Forecasting''!
Paper:
Code:
(1/3)
How does symmetry in
#NeuralNetworks
parameters impact learning? Check out our
#ICLR2024
🔥spotlight🔥 paper on ``Improving Convergence and Generalization using Parameter Symmetries''
Paper:
(1/3)
I'm not at NeurIPS this year but I strongly echo this sentiment! 🥳
My very first NeurIPS (2017, Long Beach) was also the 1st anniversary of the Deep Learning book! And I got this awesome chance to candidly chat with
@goodfellow_ian
@AaronCourville
and Yoshua Bengio.
Haha. Twitter beats me. Very excited about this work, to appear at
#NeurIPS2021
!
Existing equivariant NNs require prior knowledge of the symmetry group. We propose the Lie algebra convolutional network (L-conv) that can automatically discover symmetries from data.
(1/3)
Presenting TiDE, a time-series dense encoder for long-term time-series forecasting that enjoys the simplicity and speed of linear models while also being able to handle covariates and non-linear dependencies. Learn more →
We all know Auto-Diff is easy in
#DeepLearning
, how about automatic integration? See our
#NeurIPS2023
paper ``Automatic Integration for Spatiotemporal Neural Point Processes''.
Paper:
Code:
(1/3)
New
#ICLR2021
paper on Equivariant Continous Convolution for Trajectory Prediction. We leverage insights from
#FluidDnamics
to consider the internal symmetry in real-world trajectories.
Paper:
Code:
(1/3)
How can
#symmetry
help optimization in
#DeepLearning
? Check out our
#NeurIPS2022
paper on ``Symmetry Teleportation for Accelerated Optimization''.
Paper:
(1/3)
Scientific simulations often run at multiple fidelity. How to best combine these data in high-dimensions? See our
#ICML2023
paper on
#Deep
Bayesian Active Learning.
Paper:
Code:
(1/3)
Academic applications season has begun! So, a pic of just another normal day in La Jolla I took on my way home from a workout. 😇
#AcademicLife
#IsThisPropaganda
?🤔🤷♂️😂
How to design
#DeepLearning
method that is robust against temporal distribution shift? Check out our
#ICLR2023
paper on Koopman Neural Forecaster (KNF).
Paper:
Code:
(1/3)
How can
#GenerativeAI
help scientists discover symmetry from data? Check out our
#ICML2023
paper on ``Generative Adversarial Symmetry Discovery''.
Paper:
Code:
(1/3)
Two papers accepted in
#Neurips2019
!
- NAOMI: Non-Autoregressive Multiresolution Sequence Imputation
- Understanding the Representation Power of Graph Neural Networks in Learning Graph Topology
Thanks to amazing collaborators!
Haha. Twitter beats me. Very excited about this work, to appear at
#NeurIPS2021
!
Existing equivariant NNs require prior knowledge of the symmetry group. We propose the Lie algebra convolutional network (L-conv) that can automatically discover symmetries from data.
(1/3)
How to design Diffusion Models for spatiotemporal forecasting? Check out our
#NeurIPS2023
Paper on ``DYffusion: A Dynamics-informed Diffusion Model for Spatiotemporal Forecasting''!
Paper:
Code:
(1/3)
If you are looking for
#postdoc
positions, apply to the
@ucsd_cse
CSE Fellows Program! Join us in the beautiful La Jolla! (Email me if you work on
#ML
#AI
)
Deadline July 30, 2021.
Can I get some advice on how to deal with this
#NeurIPS
?
When all 4 reviewers give acc/strong acc for your submission, and the AC jumps in 1 day before the discussion period and pushes strongly against all reviews.
I thought paper review is a democratic process?
#NeurIPS2021
Tutorials: Shirley Ho
@cosmo_shirley
(Flatiron Institute, NYU, Princeton University) and Miles Cranmer
@MilesCranmer
(Princeton University), will be giving a tutorial on “ML for Physics and Physics for ML”.
Dec 6, 17:00-21:00 (GMT)
How to properly quantify uncertainty for multi-step sequence prediction? Check out our
#ICLR2024
paper on ``Copula Conformal prediction for Time Series Forecasting''.
Paper:
Code:
(1/3)
Interested in learning
#deeplearning
surrogate models for stochastic physics simulator? Check out our
#KDD2023
paper on ‘Deep Bayesian Active Learning for Accelerating Stochastic Simulations’!
Paper:
Code:
(1/3)
Congratulations to Dr. Rui (Ray) Wang
@Rayruw
on his awesome PhD thesis defense in
@ucsd_cse
! Ray spearheaded the area of
#Physics
-Guided
#DeepLearning
for dynamical forecasting!
Are you excited about
#AI
for
#Science
? We are hiring a postdoc on
#DeepLearning
for
#fusion
to start ASAP, in collaboration with
@GeneralAtomics
, see below for details.
It was eye-opening to visit the GA lab in San Diego. Appreciate repost!
Given all the sentiments around
#NeurIPS2020
, it might be a good time to revisit the famous experiment
@mrtz
To students:
#MachineLearning
conference acceptances are more random than we had previously realized. Let us focus on doing good science!
Join MICS faculty
@yuqirose
& Yuanyuan Shi at the launching event of the AI & Science series - Scientific ML Symposium on March 17! Details & registration at:
CSE Assistant Professor Rose Yu
@yuqirose
is the lead investigator on a new Department of Energy project to develop machine learning methods to understand climate data. Congratulations!
Read more here:
The Machine Learning for Climate conference starts today at 9 am PDT!
KITP's
#CLIMATE
-C21 conference will be in Kohn Hall's Fred Kavli Auditorium and on Zoom this week from November 1-4.
More info:
Stay tuned for recorded talks:
For people interested in physics+machine learning, check out this conference that I'm co-coordinating. . Thanks to Miles Stoudenmire for inviting me. Though I'm probably the one who knows the least about theoretical physics among others ...
I don't understand why there are so many announcements about X number of papers accepted to
#NeurIPS2020
. Don't we need to adopt reviewers/meta reviewer's suggestions, clean up and open-source the code, and prepare camera-ready?
#MachineLearning
Very proud that our superstar undergrad Peter Eckmann () was selected as a Finalist of the 2023-2024 Outstanding Undergraduate Researcher Award ( )!
Peter does amazing work on
#AI
for
#DrugDiscovery
.
I am honored to be part of the faculty cohort
@UCSDJacobs
. We met at faculty housing and have formed a strong friendship since we started. Now the friendship has turned into exciting research collaborations!
New
#KDD2021
paper on Uncertainty Quantification (UQ) in Deep
#Spatiotemporal
forecasting.
Paper:
Code:
Presentation: Aug 16 at 1 pm PST (Oral R-20)
(1/3)
I am furious to find out about Uber's
@UberEng
false claim on their
#ICML
#TimeSeries
workshop . Researchers spent a lot of effort organizing workshops. In fact, they even misspelt ICML as ''International Machine Learning Convention''...
Another example from
#Sora
of why we need
#physics
-guided
#AI
Yes. It is fun to watch soft simulations but it would be difficult to use them in high-stakes science and engineering applications.
See use cases in our DYffusion paper.
I see some vocal objections: "Sora is not learning physics, it's just manipulating pixels in 2D".
I respectfully disagree with this reductionist view. It's similar to saying "GPT-4 doesn't learn coding, it's just sampling strings". Well, what transformers do is just manipulating
For those who will have completed a Ph.D. from a U.S. institution between Jan 1, 2018, and Dec 31, 2020, and are looking for a job:
Contact me if you are into
#Physics
-Guided
#AI
,
#ML
for
#Spatiotemporal
Data.
New paper to appear at
#L4DC2022
on Neural Point Process for modeling irregularly sampled event data over space and time. (1/3)
#DeepLearning
#AI
Paper:
Code:
Come to our poster by
@claradepaolis
at
#neurips2018
relational representation learning
#R2L
workshop on neural programming DAG-to-DAG translation at 4:45 pm
New
#ICLR2021
paper on Incorporating Symmetry in Deep Dynamics model. We study symmetry in Differential Equations, applied to
#TurbulentFlow
and
#OceanCurrent
.
Paper:
Code:
(1/3)
I will be at
#KDD2023
tomorrow, giving two keynote talks
1. Uncertainty Reasoning and Quantification ()
2. Data Science for Social Good ()
In addition to our paper presentations by talented students
@allen_dongxiawu
and Brooks Niu!
We all know Auto-Diff is easy in
#DeepLearning
, how about automatic integration? See our
#NeurIPS2023
paper ``Automatic Integration for Spatiotemporal Neural Point Processes''.
Paper:
Code:
(1/3)
New paper to appear at
#L4DC2022
on Neural Point Process for modeling irregularly sampled event data over space and time. (1/3)
#DeepLearning
#AI
Paper:
Code:
Want to design
#DeepLearning
models that can generalize across dynamics? Check out our
#NeurIPS2022
paper on ``Meta-Learning Dynamics Forecasting''.
Paper:
Code:
(1/3)
I will be speaking at the upcoming workshop "Learning Collective Variables and Coarse-Grained Models ()" (April 22-26, 2024), hosted by
@IMSI_Institute
@UChicago
. (1/2)
New
#ICLR2021
paper on Equivariant Continous Convolution for Trajectory Prediction. We leverage insights from
#FluidDnamics
to consider the internal symmetry in real-world trajectories.
Paper:
Code:
(1/3)
@baaadas
I think the same question was asked 5 years ago. Look at what amazing things these PhDs are doing now! Scientific methods, critical thinking and perseverance, fundamental training in research skills goes a long way!
Our paper with
@yaguang_li
on 'Diffusion Convolutional RNN' was accepted into
#ICLR2018
! A neural sequence model that learns to forecast on a directed graph.
Come to our talk at 4:30 pm and poster at 6:30 today at
#ICML2022
#DeepLearning
-Sequence Model sessions. Learn about how to learn the right amount of
#symmetry
from turbulence data!
Want to use equivariant networks but the
#symmetry
in not perfect?
We present ``Approximately Equivariant Networks for Imperfectly Symmetric Dynamics", to appear
#ICML2022
.
Paper:
Code:
(1/3)
Come to our poster
#KDD2022
today, August 16, 4:00 PM-6:00 PM, Room 208AB (Deep Learning: New Architectures and Models) and learn about Multi-Fidelity Uncertainty Quantification! (1/3)
Paper:
Code:
Reviewing
#ML
#PhD
applications is a difficult multi-objective optimization problem. Any advice?
- publication? The randomness of
#ML
reviewing process can obscure the quality of the research.
(1/3)
Want to use equivariant networks but the
#symmetry
in not perfect?
We present ``Approximately Equivariant Networks for Imperfectly Symmetric Dynamics", to appear
#ICML2022
.
Paper:
Code:
(1/3)
Interested in
#tensors
for
#datascience
? Come by our
@kdd_news
#kdd2019
workshop on Tensor Methods for Emerging Data Science Challenges at Summit 9- Ground Level, Egan tmr Aug 5, 8am-12pm for an exciting blend of invited talks and posters!