Distinguished Professor (Emeritus), Oregon State Univ.; Former President, Assoc. for the Adv. of Artificial Intelligence; Robust AI & Comput. Sustainability
Several machine learning researchers have signed a statement regarding the upcoming launch of Nature Machine Intelligence. If you agree, I encourage you to sign this as well.
"Sentient" is being misapplied by many ML folks. It means "the ability to perceive or feel things" or "capable of experiencing things through its senses". Like many other categories, it is a matter of degree. 1/
Dear academic colleagues. It is not appropriate for class projects to be submitted to
@arxiv_org
unless they are also being submitted for publication. I and my fellow moderators are seeing more and more of this behavior. It wastes our time and the time of your colleagues
The term "ablation" is widely misused lately in ML papers. An ablation is a removal: you REMOVE some component of the system (e.g., remove batchnorm). A "sensitivity analysis" is where you VARY some component (e.g., network width).
#pedantic
Reminder to my ML colleagues.
@arxiv
is not
@github
. Don't submit drafts to arXiv; you are wasting your colleagues' time. Submit when you have a final version ready for the research community to read and cite. Please enforce this rule on your students, too.
Here is a treat: Dmitri Bertsekas is teaching RL at ASU and posting the lectures on youtube. A great way to prep for doing reinforcement learning research!
I propose that we adopt the term "Large Self-Supervised Models (LSSMs)" as a replacement for "Foundation Models" and "LLMs". "LLMs" don't capture non-linguistic data and "Foundation Models" is too grandiose. Thoughts?
@percyliang
Thoughts upon reading :
In this paper, the authors compare highly-cited papers from 2008-2009 with papers from 2018-2019 published in NeurIPS and ICML and summarize the values (i.e., desirable aspects) highlighted in those papers. 1/
@_Kitty_Wampus_
@TshepoLethea
@jasonintrator
Sprinkler systems are a hard decision for libraries. They fail surprisingly often and soak the books. This can destroy them almost as easily as fire.
@ChadNotChud
Shouldn't there be two periods: the one you are quoting and the one ending your own sentence? As in:
Nixon was lying when he said "I am not a crook.".
Disappointing article by
@GaryMarcus
. He barely addresses the accomplishments of deep learning (eg NL translation) and minimizes others (eg ImageNet with 1000 categories is small ("very finite") ?). 1/
Top 10 reasons
#deeplearning
isn’t getting us to artificial general intelligence. A critique of deep learning, 5 years into its resurgence, by
@garymarcus
I’ve been using
#AnnotatedEquations
in my recent papers. I think it really adds to the readability and understanding of the math.
Here are some examples. It uses
#tikz
in
#latex
.
Let me know if you like it. Happy for any feedback. [GitHub link: next tweet]
#AcademicChatter
+
@MelMitchell1
I also did not sign. The letter is such a mess of scary rhetoric and ineffective/non-existent policy prescriptions. There are important technical and policy issues, and many of us are working on them.
Interesting paper by Vapnik and Izmailov. "Rethinking statistical learning theory: learning using statistical invariants" shows how to impose invariants on SVM-style learning. 1/
The concept of "AGI" (a system that can match or exceed human performance across all tasks) shares all of the defects of the Turing Test. It defines "intelligence" entirely in terms of human performance. 1/
I see many papers that begin with a sentence equivalent to "Topic X is popular". Popularity is not a sound scientific reason for studying a topic, so such opening sentences strike me as lame. How about "This paper shows how to solve issue Y with method M for X"? 1/2
Nice article summarizing recent progress in deep learning. I would have titled it "Recent progress in deep learning leaves DL critics searching for new things to criticize"
I'm delighted to share the happiness of people with papers accepted to NeurIPS. But I'd be even more delighted if you write a short thread introducing your paper and telling me why I should read it. Love links to arXiv (or wherever)
Anomaly detection methods compute an anomaly score A(x), and in research, we measure their effectiveness using AUC for the binary decision "anomaly" vs. "not anomaly". But in applications, we need to choose a threshold tau. How can we set tau without having labeled anomalies? 1/
We are recruiting new people to help moderate the cs.LG (machine learning) section of
@arxiv
. If you are interested, please DM me.
Reasons to be a moderator:
1. Help promote open science and rapid communication of new results in machine learning 1/
ML Twitter: What is the current best practice for the following setting? Problem: Image classification. Setting: I'm given an initial supervised training set of labeled images drawn from P(x,y), and I train a net. Then I'm given a second set of labeled images also from P(x,y). 1/
This is a classic case of "truly-ism". It turns out every problem we solve was solvable. One day, someone will answer the core questions of intelligence and someone will say, "we thought it was difficult but it was in fact not."
"Deep learning has succeeded primarily by showing that certain questions or tasks we thought were difficult are in fact not. It has not addressed the truly difficult questions that continue to prevent us from achieving human like AI" - Judea Pearl - The Book of Why
I was hoping that the quality of COVID-19 papers submitted to cs.LG
@arxiv
would improve over time. But I think it is getting worse. Now I'm seeing more random curve fitting papers. Would anyone make life and death policy decisions based on an LSTM without any medical basis?
If this is due to machine learning, it is the most clear-cut case yet of optimizing the wrong objective. I'll bet it becomes a case study in future AI Ethics classes. Great article by
@zeynep
Gee
@garymarcus
, the goal of DL (and of the AI community) is to advance the science and engineering of intelligent systems, not to win debates or claim credit. The contributions of DL will stand or fall on their own merits. 1/
.
@rodneyabrooks
is right: the deep learning community is currently positioning itself to take credit for any future technique that anyone might come up with, without really committing to much of anything. It's a neat rhetorical trick.
I will have more to say about this, soon.
DL is essentially a new style of programming--"differentiable programming"--and the field is trying to work out the reusable constructs in this style. We have some: convolution, pooling, LSTM, GAN, VAE, memory units, routing units, etc. 8/
Very thought-provoking talk by Justin Gilmer at the
#ICML2020
UDL workshop. Adversarial examples are just a case of out-of-distribution error. There is no particular reason to defend against the nearest OOD error (i.e., L-infty adversarial example) 1/
@hardmaru
@slashML
Third, don’t try to hit a home run. Breakthroughs and insights come from surprising places, usually after you are deeply familiar with a problem. The desire to “do something big” can prevent you from seeing the insights revealed by a small, clean case
I liked this paper from NeurIPS: …
They put a deep learning wrapper around a differentiable physics engine and then can rapidly learn to fix the errors of the physics engine. They learn breakout in a few thousand steps.
@SierraClubIL
@GovPritzker
@ilenviro
I used to be a Sierra Club member. Your opposition to nuclear power is bad for the environment. Yes, there is no permanent solution, but if we don't stop CO2 emissions we know what "permanent solution" is coming.
@Meaningness
I think there are still good uses for OOP in implementing user interfaces and agent-based simulations. In such cases, it models the problem very well.
@katherine1ee
@sharongoldman
I disagree. Memorizing is the ability to correctly answer questions/info it was trained on. Generalizing is correctly answering questions it was not trained on. And hallucinating is incorrectly answering questions it was not trained on 1/
@thegautamkamath
I like to think about research as a relay race. Each paper (and its authors) seeks to hand off something useful to the next paper (and its authors). The total progress achieved is the sum of these individual contributions even if some turn out to be wrong/irrelevant
I think
@ACM_president
and
@esa
are on the wrong side of this issue. It is time to re-imagine publication models and break the stranglehold of for-profit publishers on the dissemination of scientific research. Time to rethink how research is done!
Proud to join 125+ other organizations to oppose a costly proposed Administration policy that would undermine scientific discovery, American jobs, & our global competitiveness. Read the coalition letter:
@americanpublish
@AmericanCancer
@globalIPcenter
@oneunderscore__
@emilymbender
Closing quote "If you want to run a company whose entire endeavor is to trick people into accidentally clicking on [content], then [AI] might be worth your time," she said.
"But if you want to run a media company, maybe trust your editorial staff to understand what readers want."
Prediction: In a couple of months, the US will need to do a much harsher total shutdown, which combined with testing and tracing, will do a much better job of controlling the virus. Our first try has failed to get R0 significantly below 1. We'll need a second attempt
Summary: Simple sentience (responding to sensor input) is easy to achieve, and every interactive system (including LLMs) exhibits this. LLMs mimic some behaviors that are associated with more complex forms of sentience, but there is no basis for saying they have "feelings". end/
Hey authors, when prompting ChatGPT to write your abstract, tell it you are a serious academic that does not include hype in the abstracts. I'm seeing
@arxivorg
submissions with fluent abstracts containing business-hype words. 1/
@sharongoldman
@CohereAI
@aidangomezzz
Training on synthetic data cannot lead to new knowledge discoveries. Training on synthetic data is a process of transforming one representation of knowledge into another. Any knowledge discovered by the second system must be implicit in the data generator. 1/
I'm excited to read this new (draft) book: An Introduction to Probabilistic Programming by Jan-Willem van de Meent, Brooks Paige, Hongseok Yang, Frank Wood
I've always said that Trump was intent on unilateral disarmament in international competition. Ending immigration is unconditional surrender to the rest of the world. Absolutely crazy for the US to throw away our best weapon: immigration of the best and brightest
#ICML
will have a Position Paper track. "The goal of this track is to highlight papers that stimulate (productive, civil) discussion on timely topics that need our community’s input"
#AI
Read more here:
@ReadyReporting
@mackenzief
@ndiakopoulos
@JetBlue
Perhaps even more troubling is that if you have a face that doesn't reliably match, you will be hassled every time you board. This is the repeated harm caused by poor computer vision technology.
This is a terrible idea. Machine learning is good for modeling frequent events with low stakes consequences, because it is never perfect. Nuclear launch is--we hope--an event of probability zero that requires perfect decision making.
Two US military experts have proposed giving artificial intelligence control over the nuclear launch button.
@mchorowitz
weighs in on the risks: "...training an algorithm for early warning means that you’re relying entirely on simulated data.”
Excellent post by
@AnimaAnandkumar
on requiring code release for published papers. I have not been very good about this myself, so I will make this my resolution for the New Year: The research is not done until the paper and the code are published.
Machine learning has been at the forefront of the movement for free and open access to research. For example, in 2001 the Editorial Board of the Machine Learning Journal resigned en masse to form a new zero-cost open access journal, Journal of Machine Learning Research (JMLR).
The RL community is rediscovering what folks in operations research have long known: Writing objective functions is a difficult kind of programming, and we need lots of tools and careful testing to get it right.
@isbellHFh
@Noahpinion
I'd love to see Democratic candidates run on a "Law and Order" platform just for the pure irony. Anti-corruption, anti-white terrorism, pro-rule of law.
So many important scientists (Fisher, Newton) had views that are repulsive or strange in retrospect. It makes me wonder what my ancestors believed or did. I also wonder which of my views and actions will be judged as horrifying by subsequent generations
I am pleased to announce that the camera ready version of my new textbook, "Probabilistic Machine Learning: An Introduction", is finally available from . Hardcopies will be available from MIT Press in Feb 2022.
"Frontier model" is pure hype. I encourage reviewers to insist that authors remove the phrase from their papers. What is on the "frontier" today will not be tomorrow, so it is a guarantee that your title will be wrong (probably even before you get the reviews back).
Important post from
@Noahpinion
. LLM-based tools have the potential to make all of us more efficient. Don't let the fearmongers rob us of the benefits of good AI tools.
Despite all the talk about it, work-life balance in academia remains an elusive dream.
#AAAI
rebuttal window is Friday to Sunday. As an academic I routinely have to spend weekends an evening working. This weekend,
#AAAI
rebuttals and writing
#ICLR
reviews amongst the rest. 1/3
@sbmisi
The purpose of these LLMs is not to explain jokes or to generate art but instead to learn representations and knowledge that can support many downstream tasks. The jokes and pictures are ways of assessing and demoing what has been learned.
@amanpour
@NPCollapse
Geoff
@hinton
has always been an unreliable spokesperson. When he was selling AI to Canada, he always over-sold it. He would claim we were replicating the brain. Now he has decided he doesn't like AI, so now we risk the apocalypse.
@TaliaRinger
@emilymbender
@mmitchell_ai
@ErikWhiting4
Of course, as an
@arXiv
supporter, I prefer to view arXiv as the authoritative version, and journal publication as additional evidence that the paper is worth reading. No paywalls, uniform interface, long term accessibility! :-)
@mezaoptimizer
Old time AI people like me have always been working on AGI. We just called it AI. To me AGI is just a marketing term. We old timers think we know how hard the problem is, and we are lolling as the younger generation discovers this
Every time we read a great paper, we should tweet or blog about it. This would be particularly valuable for researchers who don't have a PR department to trumpet their accomplishments. Much more meaningful than bibliometrics. 2/2
@KLdivergence
I think the biggest difference is that ML people are trying to build software systems, whereas statisticians seek to support scientific inquiry. This is why ML folks don't typically care about estimation or statistical inference
@erikbryn
The technique in this paper is more suited to algorithm discovery rather than scientific discovery. It relies on having a method for verifying a proposed solution. The LLM doesn't know the answer, but it can generate good proposals. 1/
@sapinker
The Board wasn’t “forced” to change the name. We welcome the new name and the name change is only one of many steps we are taking to make our community more welcoming to everyone.
Wonderful blog post by
@yoavgo
on "Reinforcement Learning for Language Models". I especially like the insight that training for "truthfulness" depends on the teacher knowing what the LLM believes.
To researchers submitted Covid-19 lung x-ray papers to
@arxiv_org
: If you test on a small test set (e.g., 100-200 images), you can't possibly measure error rates, AUC, etc. to 4 significant digits. Nobody will take you seriously if you don't compute confidence intervals!
I think we should be building systems that complement people; systems that do well the things that people do poorly; systems that make individuals and organizations more effective and more humane. 3/
Anecdotally, I find that if I run speedtest at the same time as a file download, the file downloads much faster. Is
@comcast
@xfinity
detecting speedtest and adapting?
I am recruiting PhD students interested in how AI systems can detect and respond to surprise. Methods include anomaly detection, uncertainty assessment, calibration, and model-based RL. If you are at
@icmlconf
, DM me and we can chat
I agree that existing ML/AI systems focus on closed worlds. This is the fundamental reason that these systems are not safe to deploy in high-stakes open-world applications. But the idea that knowledge engineering will avoid these problems is puzzling. 1/
I will be at
@NeurIPSConf
and would be eager to meet with people. I'm not recruiting or offering big signing bonuses, I'm just interested in learning about what people are working on and what difficulties they are encountering. DM me and we can schedule a time to chat
@rao2z
@ylecun
But stepping back a bit, are we formulating the problem correctly? Our algorithms are measuring expected error, but we want to minimize something like worst-case error over the entire input space. Relying on expected error inevitably produces bias
@MuslimIQ
"Tear gas is a chemical weapon banned in war. But Ferguson police shoot it at protesters." The US is banned from using tear gas on the battlefield, but we use it on our own citizens routinely
The pre-training of LLMs ignores "factuality". Has anyone developed loss functions for optimizing factuality? I'd be interested in a philosophical analysis, too. A related question: What is the best place to read about automated fact checking? 1/
It is also strange to speak of LLMs as being sentient when the only sensor they have is the incoming text stream. Certainly humans can experience social pain when exchanging text message (as twitter proves every day), but LLMs are not social agents 11/