Thrilled to release our NeurIPS '21 Spotlight, led by students Rohan Mukherjee and Yeming Wen: "Neural Program Generation Modulo Static Analysis". . tl;dr: In program synthesis and related tasks, Codex/GPT is not all you need. (1/n)
"We find that the effort researchers waste in writing proposals may be comparable to the total scientific value of the research that the funding supports, especially when only a few proposals can be funded."
This paper () has now been accepted at
#TOPLAS
and will appear at
#PLDI24
as a journal-first presentation. Proud of
@meghana_aparna
's excellent work on this, and many thanks to my amazing collaborator Tom Reps!
Let me now summarize why I am excited about
Some CS papers are written over a few months -- some others, over a year. Our new Arxiv post, led by Tom Reps and
@meghana_aparna
, describes ideas developed over ~25 years.
(1/n)🧵
It's official: I am moving to
@UTCompSci
in Spring 2020!
@RiceCompSci
is a wonderful department -- I owe a tremendous amount to my students, mentors, and collaborators here -- and Houston has a lot going for it too. But after eight years, it was time for a change.
Some CS papers are written over a few months -- some others, over a year. Our new Arxiv post, led by Tom Reps and
@meghana_aparna
, describes ideas developed over ~25 years.
(1/n)🧵
Program induction is hard as search spaces of programs explode quickly. We have a new line of attack on this (to appear at
#NeurIPS2020
): using neural relaxations of discrete sets of programs as admissible heuristics. Paper: . Code: .
Delighted to announce PutnamBench, a new AI-for-math benchmark for evaluating neural theorem provers for Coq, Lean, and Isabelle on Putnam math competition problems. Almost all the problems here are beyond the reach of current approaches. Excellent leadership by
@gtsoukal
, who
Announcing PutnamBench: an evaluation benchmark for formal mathematical reasoning in Lean 4, Isabelle, and Coq! PutnamBench consists of problems from the William-Lowell Putnam Mathematical Competition, the premier collegiate mathematics exam in the US & Canada. 🧵
We had a lot of fun writing this survey! Thank you,
@rupakmajumdar
, for inviting us to write this.
You can find a free excerpt here:
Below, a 🧵on what neurosymbolic programming (NSP) is and why we think it's important. (1/n)
Hey, ML/PL enthusiasts! Looking for some "light" reading for the holiday break?
FnT just published our survey on "Neurosymbolic Programming", written jointly with
@swarat
, Kevin Ellis,
@rishabhs
, Armando Solar-Lezama, and
@yisongyue
.
1) This is a thread on our
#NIPS2018
paper (
@lazarvalkov
et al.; preliminary version at ), and more broadly, about the role of language abstractions in deep learning (DL).
A review in Nature, by
@candice_odgers
, asserts that I have mistaken correlation for causation and that “there is no evidence that using these platforms is rewiring children’s brains or driving an epidemic of mental illness.” Both of these assertions are untrue.
Our Neurosymbolic Programming tutorial is coming to
#POPL23
! We'll explain the basics, do an algorithmic deep dive, and explore neuroscience applications.
* 1/16 (Mon), 2 pm
* Speakers: Armando, my student Atharva, me (
@yisongyue
&
@JenJSun
in spirit).
The
@neurosym
summer school is off to a great start!
In the morning,
@atharva_sehgal
,
@akavidemic
, and I gave the first part of our tutorial on neurosymbolic programming. Slides at .
Now
@AI4Code
is telling us about the Scallop framework for
The neurosymbolic learning meetup at
#ICLR2024
was a big success! Many thanks to all who showed up — and especially to
@theo_olausson
, who made the event happen.
With
@mvechev
and Armando in Vienna. We first met 18 summers ago, as grad student interns at IBM research. In the years since, we have grown up, picked up new research tools, acquired new responsibilities. But when we see each other, it’s 2006 again.
I am beyond excited to be part of this new
@NSF
CISE
#Expedition
on AI for systems: .
Our goal is to build a new kind of OS in which much of the decision-making is done by ML. This is a perfect playground for research on trustworthy/verified ML and
Absolutely thrilled about this landmark achievement, both as a computer scientist and as someone about to spend a year with the
@GoogleDeepMind
team behind this work.
Solving International Maths Olympiad (IMO) level problems has been a grand challenge for AI systems. Happy to share that a new solver (AlphaProof) developed by our team
@GoogleDeepMind
and our geometry solver (AlphaGeometry) were able to solve 4 out the 6 IMO 2024 problems!
Excited to be a part of this venture! 🚀
@AsariAILabs
() is hiring, and if you are in the job market and excited about AI for difficult, real-world engineering applications, you should talk to us.
What makes Asari different? LLMs have excelled at low-level
A new journey begins – 🚀 we’re excited to launch!
Our mission is to build AI that helps us co-invent the future.
-- And we’re hiring to make this happen👇
We’re building a new type of AI agent and tools that help us imagine and create 10x better solutions, products, and
Skipping
#PLDI
for the first time in a while, but there's a cute reason why. Ateesh Peterson Chaudhuri, joint project with the amazing
@TL_Peterson
, arrived on June 13. Life is good (though full of billions of poopy diapers)!
PL/Formal methods twitter: what are some examples of academic ideas from PL/FM from the last 25 years that have had demonstrable real-world impact? To start the list:
1) SAT-based model checking, applied to hardware.
In just 11 years,
#ICLR
has become one of the most innovative and exciting events in CS. I am thrilled to help
@yisongyue
,
@_beenkim
, and the other Program Chairs run the 2024 edition in beautiful Vienna. If you have suggestions/comments, please send them our way!
I am honored to serve as Senior PC for
#ICLR2024
. Looking forward to serving with
@_beenkim
(General Chair) and a fantastic PC cast (
@YizhouSun
,
@swarat
,
@EmtiyazKhan
, & Katerina Fragkiadaki). Hope to see you all in Vienna!
Paper on "neurosymbolic" program synthesis for lifelong learning, authored with
@RandomlyWalking
and students, accepted at
#NIPS2018
! Moral: functional idioms and type-directed synthesis can facilitate transfer across learning tasks. Preliminary version at
Many congrats to Abhinav Verma (), who aced his PhD defense today. Abhinav worked on a new kind of RL based on neurosymbolic program synthesis. He will start as a prof at Penn State after a prebbatical with
@thenzinger
. Read his papers, and work with him!
How do you learn neural networks that respect end-to-end safety requirements of larger systems of which they are a part? Our new ICLR paper, led by
@ChenxiYang001
(), explores this question. (1/n)
Important post by
@GaryMarcus
on the limitations of DL. PL/logic researchers take note; we have much to contribute to this debate. In the recent past, we've had LEAPS in theorem proving, program synthesis, etc. We should try to leverage these ideas (+ DL) in classical AI tasks.
United Airlines canceled my flight, British Airways bumped me off a flight, American Airlines lost my bag. But I am finally in beautiful Vienna, on time for the first day of
#ICLR2024
, looking forward to:
* The exciting announcements that
@yisongyue
will make in his opening
Today's highlights:
1) The baby possibly has an ear infection
2) The toddler is going stir crazy and repeating his demands slowly and loudly when told no
3) A possum family has moved into our attic
4) Too much time spent on Twitter.
No, I haven't figured out Zoom classes yet.
The
#ICML2024
AI-for-math workshop starts in 10 minutes! If you are here in Vienna, consider stopping by.
I will give a talk at 9:35 am on sequential-decision making agents for mathematical discovery. I'll post the slides here right after the talk.
Many fun announcements about LLM agents in the last few weeks! I'll add one from our lab: Copra, a retrieval-augmented GPT-4 agent for formal theorem-proving in frameworks like Coq and Lean.
Copra, developed by
@AmitayushThakur
in collaboration with George Tsoukalas,
@YemingW
,
Our
@NSF
Expeditions project "Understanding the World with Code" gets kicked off next week (Oct. 5-6)! We aim to build up a science of neurosymbolic programming and use it to make new natural-science discoveries. . Livestream: . [1/2]
The website for our NSF CISE Expedition on "Understanding the World with Code" is now up: . We are eager to connect with other folks working on AI/ML and PL for the sciences. If you are interested, please reach out!
I can't say enough about how excited I am about this Expedition! I think a program synthesis perspective can really help AI approaches to the natural sciences. Our interdisciplinary team, led by Armando Solar-Lezama, will show how. More context here:
Announcing the 1st Austin Workshop on Program Synthesis, a.k.a. the final project poster session (virtual) for my class CS 395T. Feel free to stop by tomorrow between 3-5:30 CT and check out a poster or two!
PSA:
@adityaakella
and I are looking to hire a postdoc for an exciting new project at the interface of formal methods, machine learning, and software systems. Our high-level goal is to build scalable learning-enabled systems with strong reliability guarantees. A good candidate
Nice
@CACMmag
article by
@donmonroe
on neurosymbolic learning. It's been wonderful to work in this area over the last few years -- there are so many open problems and new applications! Increasingly, we are seeing a convergence... (1/3)
Really looking forward to this visit to the alma mater. It's always exciting -- and just a little bit intimidating! -- to speak at the room where you defended your Ph.D. thesis.
Many fun announcements about LLM agents in the last few weeks! I'll add one from our lab: Copra, a retrieval-augmented GPT-4 agent for formal theorem-proving in frameworks like Coq and Lean.
Copra, developed by
@AmitayushThakur
in collaboration with George Tsoukalas,
@YemingW
,
Looking forward to today's workshop at
#ICML2023
! I'll talk about
@chenxiyang_ut
's work on formally certified learning. Our goal: train agents that mix human code and differentiable NNs and can invoke verifiers as tools during learning. Room 310, 10:40 am.
[1/9] We are looking forward to seeing you all tomorrow at the Differentiable Almost Everything workshop
#ICML2023
We will start at 9am (Hawaii time) in Room 310.
I am honored to be a member of this year's cohort of
@TheOpEdProject
Public Voices Fellowship. My first op-ed, on risks from personalized AIs that turn into AI "frenemies", appears in
@thehill
today.
Thrilled about this work led by my Ph.D. student
@YemingW
, now accepted as an oral presentation at
#ICLR2024
!
The work takes on a basic issue with LoRA: that it can't efficiently serve multiple domain-specific adapters at the same time. The solution is a new batching mechanism.
🚀 Excited to share that our paper "Batched Low-Rank Adaptation of Foundation Models ()" was accepted for an oral presentation at
#ICLR2024
! We're advancing parameter efficient fine-tuning (PEFT) in LLMs for more personalized and efficient AI. (1/n)
Last week's
@neurosym
summer school was a blast! Many thanks to all who attended and presented. Here's a group photo in front of the beautiful Salem harbor—thank you,
@konet
, for taking it!
All talk slides will be available at .
Great start to Day 3 of the
@neurosym
summer school!
@ZennaTavares
is telling us about the ChiRo system for learning and causal reasoning that his team is building at
@BasisOrg
.
Inspiring talk by
@RanjitJhala
on language-integrated verification at
#pldi18
. If you like PL/FM research and weren’t here, you need to check out the videorecording later.
"Programmatically interpretable reinforcement learning" Verma et al.,
#themorningpaper
RL policies that are human interpretable and verifiable - i.e., deployable!
Want to do a postdoc on either foundational or applied machine learning with a world-class group? Consider applying to our NSF AI Institute. If accepted, you can work with anyone in our large team; ML + PL/formal methods fully in scope. Deadline Dec. 15.
A few months ago, I had an enjoyable conversation with
@MitchWaldrop
on neurosymbolic learning. His article on this topic at PNAS Front Matter is now available, and it's right on point. 🧵
Had so much fun teaching my new class on "Logic in CS and AI" today! The first lecture follows
@cdixon
's brilliant 2017 article (), starts with Aristotle, and ends with verified systems and neurosymbolic programming. Slides here:
Neurosymbolic ML is a natural fit for natural science applications -- our new position paper gives a detailed argument as to why. The paper was led by
@konet
,
@JenJSun
, Megan Tjandrasuwita, and Atharva Sehgal.
I am very excited to share with all of you my first research paper as a senior researcher
@MIT_CSAIL
! “Neurosymbolic programming for science” Here we introduce the opportunities of neurosymbolic programming techniques to accelerate scientific discovery:
A very nice set of blog posts on the
#ICML2024
#AIforMath
workshop by Harald Carlens of
@ml_contests
. If you are excited by Alphaproof and want to know what's going on in the broader field, read these posts.
* Morning session:
* Afternoon session:
Chuchu's thesis is an example of what forward-looking formal methods research looks like. Chuchu shows that you can effectively combine white-box verification with more general data-driven methods, tremendously boosting the scope of FM. Congrats again on a well-deserved award!
Congratulations to Chuchu Fan, whose dissertation received the 2020 ACM Doctoral Dissertation Award for making foundational contributions to verification of embedded and cyber-physical systems. Happy
#AdaLovelaceDay
to all
#WomenInSTEM
!
It's been a sad couple of days at UTCS.
I last saw Will Cook at a dinner a month or so ago. We talked about meeting up for a drink and chatting about research and life. I should have followed up.
The costs of LLM inference have gotten much attention lately, and rightly so. Over the last year, we have been thinking about ways to reduce these costs by compiling LLM queries into queries for "programs" over LLMs and smaller, cheaper models. Our first effort on this topic,
🚀 Thrilled to present our paper "Online Cascade Learning for Efficient Inference over Streams," to appear at
#ICML2024
! 🎉 We've crafted a new way to switch between LLMs and cheaper models learned online, significantly reducing inference costs.
Great start to Day 3 of the
@neurosym
summer school!
@ZennaTavares
is telling us about the ChiRo system for learning and causal reasoning that his team is building at
@BasisOrg
.
Many congratulations to Dr. (and soon, Doctor-Professor) Greg Anderson!
Greg did serious, interesting work coupling formal methods and ML, especially deep RL. Here's a 🧵summarizing some of his results. I think he will be an amazing asset for
@ReedCollege
! (1/4)
So proud of the grad students at
@UTCompSci
for launching this program! If you are from an underrepresented group and are applying to CS PhD programs this year, please sign up. Our student mentors will offer you quality feedback on your application.
A new student-led initiative from the graduate student group (GRACS)
@UTCompSci
to help under-mentored PhD applicants with feedback on their application material.
Send in your application material before Nov. 27th.
Are you a PhD student or advanced undergrad interested in neurosymbolic learning? Apply to attend the second
@neurosym
summer school!
The school will be held in beautiful Salem, MA, from June 10-12. We have a great speaker lineup, and you will likely meet a lot of students who
Join us for our second
@neurosym
summer school 2024 to be held at Salem, Massachusetts June 10-12 a historic and cultural town near Boston. Applications due May 6th!
An excellent first day of a summer at Facebook HQ. Here to aid FB's growing efforts on AI for code. Programming language abstractions + machine learning FTW!
A year ago, I was stressing about the coming LLM-induced end of open ML+X research. Tom Reps, one of the wisest people I know, told me to relax -- he had seen this panic before with Windows and Linux. I'm increasingly convinced that Tom was right. Open-source finds a way.
Companies trying to raise and invest 100s of millions or billions of dollars training foundation models should be worried right now - a big twist in the AGI story is incubating. Current setups are akin to first mainframe computers - big and clunky. Will be interesting to see.
Can you have deep RL with formally verified exploration/policy update steps? Seems hard; NN verification is costly! Our new work gives a path forward: mirror descent + shield synthesis enable verified deep RL sans direct NN verification.
#NeurIPS2020
[1/2]
If you are attending
#PLDI2018
, come to our June 18 (Monday) tutorial on deep learning for program synthesis /analysis! No background in synthesis or ML necessary.
Finished a fun two months of work at Facebook HQ on statistical program synthesis. Very impressed by the energy in the Big Code team; some quite promising results. More on these in a few months, hopefully!
Just finished reading the Voyager paper () -- thrilled to see programmatic representations and library learning scale to tasks this sophisticated, and amazed that GPT-4 can do so well at not just coding but also defining coding tasks. (1/n)
RL is too difficult to learn from scratch. The future is large foundation models with world knowledge helping with automatic curriculum generation and online learning
An exciting opportunity for those about to finish their PhDs! I would be delighted to host postdocs interested in the intersection of PL/logic and machine learning, including program synthesis, probabilistic and differentiable programming, and safe autonomy.
I really enjoyed my visit to
@INSAITinstitute
in Sofia last week. I don't think I've ever gotten this many questions at a talk! INSAIT, led by
@mvechev
, is an amazing effort that can transform the CS research landscape of Eastern Europe. I look forward to watching it grow.
Last Thursday was the first lecture of the new tech series of INSAIT, where top scientists who come up with the latest innovations in AI and computing talk about them in Sofia. Thanks for the amazing lecture by Prof. Swarat Chaudhuri (
@swarat
) on Neurosymbolic AI.
Abhinav Verma's paper "Programmatically Interpretable Reinforcement Learning" accepted at
#ICML2018
! A decent example, IMO, of how PL ideas can aid the quest for safe and accountable AI. So honored to work with such inspiring students and collaborators!
Our NIPS 2018 paper on FP4DL, now named "Houdini: Lifelong Learning as Program Synthesis", is out! . Houdini stands for either "Heuristic Optimization for the Ultimate Development of Integrated Neurosymbolic Intelligence", or just Hou(ston) + (E)din(burgh).
My mentee Josh Michalenko defended his MS thesis today. His topic: interpreting the state space of RNNs trained to recognize regular languages. Finding: such an RNN represents an “abstraction” of the minimal DFA for its language. See our
#iclr2019
paper:
Thank you for profiling us (again),
@adriancolyer
! Here is a different example of the use of functional abstractions in program synthesis, by
@polikarn
et al.:
I am thrilled to be part of this
#ICLR2024
paper with
@lazarvalkov
,
@RandomlyWalking
, and
@variational_i
! The paper presents a probabilistic search technique for scaling up modular continual learning, following up on our earlier Houdini framework for lifelong learning through
I’ll be presenting our
#ICLR2024
paper on a probabilistic approach to scaling modular continual learning algorithms while achieving different types of knowledge transfer. (, in collaboration with
@variational_i
@swarat
@RandomlyWalking
). A tldr (1/8):
FINAL UPDATE: On June 24th, Armando Solar-Lezama (Professor in EECS and COO/Associate Director of CSAIL, MIT), Tonio Buonassisi (Professor of Mechanical Engineering, MIT), and Yoon Kim (Assistant Professor in EECS and CSAIL, MIT) released a public statement regarding the paper.
The
#ICML2024
AI-for-math workshop starts in 10 minutes! If you are here in Vienna, consider stopping by.
I will give a talk at 9:35 am on sequential-decision making agents for mathematical discovery. I'll post the slides here right after the talk.
If you are at
#ICLR2024
, check out
@atharva_sehgal
's nice new work -- with Arya Grayeli,
@JenJSun
, and me -- on compositional world modeling!
The method introduces a new neurosymbolic representation of entities in a changing world. It uses the compositionality of symbolic
Excited to present Neurosymbolic Grounding for Compositional World Modeling at
@iclr_conf
!
Our neurosymbolic algorithm learns world models from unsupervised interactions in a novel compositional generalization environment.
More info here:
#ICLR2024
Come to the first summer school on neurosymbolic programming! We have an exciting lineup of speakers and tutorials on symbolic + neurosymbolic program synthesis and probabilistic programming. You can apply at .
New paper on learning programmatic policies accepted at
#NeurIPS19
! tl;dr: We cast program synthesis as a form of projected gradient descent where one alternates between gradient steps in a neural functional space and projections into a programmatic space.
I had a lot of fun writing this post on the synthesis of neurosymbolic programs, the technical challenge at the heart of our new NSF Expedition (). Thank you,
@michael_w_hicks
, for being an amazing editor!
Today on PL Perspectives:
@swarat
introduces *neurosymbolic programs*, which are constructed by combining neural network training and PL-style program synthesis.
At lovely Merida, Yucatan, to give my first-ever robotics conference talk at
#WAFR2018
: . Every talk here has to have a "dirty laundry" slide that describes the work's limitations. This is an excellent idea -- all research communities should adopt this!
Many congratulations to Dr. (and soon, Doctor-Professor) Greg Anderson!
Greg did serious, interesting work coupling formal methods and ML, especially deep RL. Here's a 🧵summarizing some of his results. I think he will be an amazing asset for
@ReedCollege
! (1/4)
Delighted to introduce Dr. Anderson! It's been a pleasure to work with Greg and
@swarat
on Safe Exploration for Reinforcement Learning and see him through to his graduation. Looking forward to seeing what he will do at Reed College!
Very bummed to miss
#NeurIPS2023
, but several of my students are attending. Please say hi to them if you are interested in our lab's work!
-
@AmitayushThakur
and
@GTsoukalasRU
will be at the
#MATHAI
workshop, where we present our work on LLM agents for Lean/Coq theorem-proving:
Now accepted at
#NeurIPS2022
! Joint work with Cameron Voloshin (lead),
@HoangMinhLe
, and
@yisongyue
.
tl;dr: We introduce the problem of policy optimization under LTL constraints and give a solution with a rigorous sample complexity analysis.
Preprint:
Thrilled to be part of this effort! A big thank-you to leader Cameron Voloshin and co-authors
@yisongyue
and
@HoangMinhLe
.
LTL has long been a cornerstone of formal methods. Recently, LTL has found another use: as a language for communicating human intent to autonomy. (1/7)
Thrilled to be part of this effort! A big thank-you to leader Cameron Voloshin and co-authors
@yisongyue
and
@HoangMinhLe
.
LTL has long been a cornerstone of formal methods. Recently, LTL has found another use: as a language for communicating human intent to autonomy. (1/7)
Policy Optimization with Linear Temporal Logic Constraints:
1st author: Cameron Voloshin
Co-authors:
@swarat
&
@HoangMinhLe
LTL can capture expressive constraints that are hard to do with reward engineering, such as an infinite loop (e.g. patrolling).
Episode 43 of Thesis Review:
Swarat Chaudhuri (
@swarat
), "Logics and Algorithms for Software Model Checking"
We discuss reasoning about programs, formal methods & safe machine learning, and the future of program synthesis & neurosymbolic programming.