Robocrunch

Yann, yes it is true that the best current systems for image recognition, etc. don't use explicit probabilistic inference, but I would argue that all serious, deployed current systems for embodied AI (localisation, mapping, SLAM, 3D scene understanding) rely strongly on it. 1/n

Shared by Andrew Davison at 5/14/2022

Important ingredient for long-term productivity: having a reliable activity to clear your mind. For me it's running or weightlifting. For first half hour, I am still thinking about the problem of the day; by an hour I'm a blank slate. Hard problems somehow feel far simpler after.

Shared by Greg Brockman at 5/14/2022

Here is a recipe to debug a hanging python multiprocess setup (e.g. pytorch distributed). To get all tracebacks: pip install py-spy ps aux | grep python | grep -v grep | grep `whoami` | awk '{print $2}' | xargs -I {} py-spy dump --pid {} you may need to add `sudo` before py-spy

Shared by Stas Bekman at 5/14/2022

Artisanal architectures, small batched and locally sourced

Shared by Brandon Rohrer at 5/14/2022

10 Terminal Commands Anyone Learning Python Should Know by @frankandradec https://t.co/XlalksSMw8

Shared by Towards Data Science at 5/14/2022

“Australia’s electoral system, in contrast, is praised by analysts around the world … the ‘gold standard in election administration’. Few other countries have independent electoral commissions. ‘People in Australia trust those processes’.” https://t.co/MM8OU48o6F

Shared by Christopher Manning at 5/14/2022

This is not a statement against modeling uncertainty. I'm all for that! (Though, I don't think *probabilistic* models are necessary for that). It's a statement about having given too much importance to it *at the expense* of other key questions, like representation learning.

Shared by Yann LeCun at 5/14/2022

The EMNLP 2022 logo features an eight-pointed star constructed from rotated repetitions of the ACL’s iconic logo. To learn more visit the EMNLP2022 Blog. https://t.co/k4thsYJNpQ

Shared by EMNLP 2022 at 5/14/2022

Second post is up! Judge Judy, causality, composer, and a bonus! https://t.co/s7oFgbcqf0

Shared by Andrew Carr at 5/14/2022

VIDEO 🎥 A legacy of legal uncertainty continues to mar any potential agreement on data flows between the U.S. and EU. Here's why it's so important to get this deal right: https://t.co/6fpwXa8u7p

Shared by Data Innovation at 5/14/2022

Happy that my friend @ilyasut has been elected a Fellow of the Royal Society! As he has also been gifting us some interesting tweets lately, I have challenged a Language Model 🐹 to see if it can compete. Results in thread🧵(1/8)

Shared by Oriol Vinyals at 5/14/2022

In the age of increasingly ubiquitous and powerful AI, your datasets are your moat.

Shared by Bojan Tunguz at 5/14/2022

The Curse of Delayed Performance - Predict the performance of your model - before the ground truth is available. @TheRealNannyML https://t.co/cHyHJ7z4oR

Shared by KDnuggets at 5/14/2022

Robot uses high-power lasers & AI to instantly identify and target weeds to eradicate them. It doesn't disturb the soil or use any herbicides. credit: @pascal_bornet @carbonrobotics

Shared by MIT CSAIL at 5/14/2022

I fear that AI research is being increasingly absorbed into the second category.

Shared by François Chollet at 5/14/2022

Hi Murray. Gato is tiny. It needs development. With Lamda, Codex, AlphaCode, AlphaFold, Flamingo, etc, the evidence is strong for the hypothesis that scale is going to take us to AGI, but by scaling I mean addressing the challenges in my tweet, not just more parameters.

Shared by Nando de Freitas at 5/14/2022

"A Comprehensive Survey of Image Augmentation Techniques for Deep Learning" https://t.co/AJ89b0Ggka --Overviews like this are actually super helpful. Would be a nice future project to build sth like "model tables" where in addition to image aug. we also list hyperparam settings

Shared by Sebastian Raschka at 5/14/2022

“Causal” is like “error term”: it’s what we say when we’re not trying to model the process. https://t.co/29sv2iAEFj

Shared by Andrew Gelman et al. at 5/14/2022

Our next Revolutionizing Healthcare session is in the making! Clinicians can join us on 25 May when we discuss the highly relevant topic of 'using machine learning to power clinical trials' at an engaging roundtable with some outstanding experts. Sign up: https://t.co/qniCY99WwL

Shared by Mihaela van der Schaar at 5/14/2022

Join me 🤩 and @JotyShafiq in congratulating 🎓Dr Samson Tan @samsontmr 👨‍🎓on his dissertation defense of his thesis✨Linguistically-Inclusive NLP✨! 👏👏 Thanks to Hwee Tou Ng, Michel Galley, @dirk_hovy, Seth Gilbert & Roger Zimmerman w the defense! Now w @awscloud AIRE! ☁️✈️ htt

Shared by Min-Yen Kan at 5/14/2022

An interesting paper which proposes an ML-based approach to rectify situations in which the smoothness assumption of the manifold hypothesis does not hold e.g., data with singularities. 1/n Algebraic machine learning with an application to chemistry: https://t.co/Vh85mWyU5i

Shared by Stephen Ra at 5/14/2022

Interesting thread about the future of large models, economic implications and the path to commercialization. It’s still a question mark whether startups can build a profitable and sustainable business model around large models, or whether these models simply become a commodity.

Shared by hardmaru at 5/14/2022

Order varies, but 2016-2021 news is best represented in the Common Crawl in these languages: Eng,Spnish,Frnch,Grmn, Rssian,Itlian,Portu,Hindi,Trkish,Arbic,Romnian,Dutch,Swdish,Krean,Jpanese,Vitnmese,Chinse,Greek,Indnesian,Ukrinian,Plish,Czch,Hngarian,Blgarian,Prsian,Norwgian,Thai

Shared by Unso Eun Seo Jo at 5/14/2022

It makes sense. The precision in BF16 is abysmal - you almost want to do stochastic rounding to make it unbiased! We know it's better to accumulate at higher precision, and during training the weights are kinda getting accumulated.

Shared by Leo Dirac at 5/14/2022

And here you were thinking your human vision system wasn’t prone to attacks

Shared by Reza Zadeh at 5/13/2022

"Scaling-Translation-Equivariant Networks with Decomposed Convolutional Filters", by Wei Zhu, Qiang Qiu, Robert Calderbank, Guillermo Sapiro, Xiuyuan Cheng. https://t.co/t5t3A27x7i

Shared by Journal of Machine Learning Research at 5/13/2022

AMTA (Association for Machine Translation in the Americas) call for papers is out - conference deadline is on June 6! Conference will be held on September 12-16 and will be hybrid, with the in person part in Orlando, FL https://t.co/4nbanz9gB3 #NLProc #neuralMT #neuralempty

Shared by Huda Khayrallah at 5/13/2022

Models will be able to produce satisfying outputs for a given reasonable prompt as robustly as humans do. Being able to deal with various prompts as flexibly as humans do is a part of general intelligence and unavoidable. https://t.co/HnlcPiM3nu

Shared by Aran Komatsuzaki at 5/13/2022

if a career in crypto isn't high-risk enough for you and you'd like to join a humanoid robotics startup, hmu! I'm looking to grow my team soon

Shared by Eric Jang at 5/13/2022

The official Swin-V2 impl (https://t.co/RxRQ6OtvUD) did have some differences from the timm 'CR' impl, but the core of it was the same and a merged model that supports both variants wouldn't be too hard. The MLP Log-CPB had most diff w/ normalization + sigmoid but both work.

Shared by Ross Wightman at 5/13/2022

How could we build AI systems to reuse existing models rather than learn a completely new model each time? This is the question @TusharKhot, Kyle Richardson, @DanielKhashabi, and Ashish Sabharwal seek to answer in their papers, described in 2 parts: https://t.co/1n4DqrJIkk

Shared by Allen Institute for AI at 5/13/2022

Users can do all of this through our Python SDk, the UI and through PQL (Predictive Query language), an extension of SQL that brings ML and data together in a familiar language.

Shared by Piero Molino at 5/13/2022

Delighted to see @GoogleAI collaborate w/@MGnifyDB to use ML to provide 1.5 billion protein function predictions (!). While protein structure prediction is useful, accurate protein function predictions can truly unlock all kinds of new possibilities!

Shared by Jeff Dean (@🏡) at 5/13/2022

Join us today 4 PM ET to hear @summerfieldlab talk about human and machine intelligence at the #LearningSalon with me @csuncodes & @blamlab! https://t.co/19nEdZUnAG

Shared by Ida Momennejad at 5/13/2022

#NVIDIAInception startup, WeRide, launched its fleet of 50 driverless vehicles performing round-the-clock cleaning services. Learn how the company is currently building next-generation self-driving solutions on NVIDIA DRIVE Orin. https://t.co/lEIrdFhZbl

Shared by NVIDIA AI at 5/13/2022

Several AI documentation processes have been developed in recent years to improve transparency and accountability. Discuss lessons learned and new ideas on May 25th with @sociotiose, Emily McReynolds, @JessicaH_Newman, and PAI's @ChristineCustis. https://t.co/mHD3TNcr8a

Shared by Partnership on AI at 5/13/2022

EntailmentBank, our corpus of detailed tree-structured explanations, is now available visually as a 700 pg book (PDF), making brainstorming your next explanation-generating algorithm much easier. @allen_ai PDF downloadable here: https://t.co/1jo8CQIMcb

Shared by Peter Jansen at 5/13/2022

who can land on the moon, race, climb mountains, bipedal and more, all with ML + reinforcement learning? https://t.co/N7iFOhmTLx

Shared by clem 🤗 at 5/13/2022

Nice Article "Data Eats the World" https://t.co/q89RN2Ndin #KnowledgeGraph #GraphDB #Reasoning #TensorFlow

Shared by bernhard at 5/13/2022

Using machine learning to add 1.5 *billion* protein function predictions to @MGnifyDB - collaboration between @emblebi and @GoogleAI !

Shared by Zoubin Ghahramani at 5/13/2022

The data on which the big transformer was trained was generated from independently trained agents, was it not? There is no actual RL being done to train the final agent, it's all amortizing other agents, and each of the RL-trained agents was trained on a single environment.

Shared by David Pfau at 5/13/2022

Leaving the #newnlp workshop inspired by all the great people working on new languages for NLP in DH — keep an eye on them! And thanks to @PrincetonDH for helping create a space for all this impactful work to happen. https://t.co/WRrQmHZGJ9

Shared by David Bamman at 5/13/2022

[Meetup] The Present and Future of NLP with Jacob Devlin, creator of the BERT algorithm 📣 https://t.co/SOaYLZyxnw

Shared by NLPlanet at 5/13/2022

Fundamentally, how do you determine what a "spam bot" or "fake account" actually is, without machine-computable terminology (a/k/a an #Ontology) ? We all know that one person's networking is another person's spam! Building networked apps without #Identity #Authenticity is DOA++

Shared by Kingsley Uyi Idehen at 5/13/2022

This paper is a great example of where linguistics and morphology can shine ✨ in NLP—in languages with different linguistic properties and low-resource settings.

Shared by Sebastian Ruder at 5/13/2022

Did you know that differentiating a volume renderer will produce biased and noisy derivatives? Our new sampling technique fixes this, improving reconstruction of editable & relightable volumes. Joint work with @merlin_ND, Thomas Müller and Alex Keller at SIGGRAPH'22. (1/8)

Shared by Wenzel Jakob at 5/13/2022

Learn how our developer community solves real, everyday machine learning problems with PyTorch. From Advertising & Marketing to Travel and so much in between, get to know PyTorch’s features and capabilities. Read all about PyTorch’s Community Stories: https://t.co/ceOqYIL5fR

Shared by PyTorch at 5/13/2022

Update for those who sent me a request to be a NeurIPS reviewer: I was too late to put you in for NeurIPS, but I now have your name in the list, so I hope I can nominate in the future. For now, I have added about 25 or so reviewers for ACML 2022, and I hope you will accept.

Shared by Emtiyaz Khan at 5/13/2022

Single-cell technologies can map human cellular systems at unprecedented scale and resolution. New research in @ScienceMagazine uses this cutting-edge technology to create a comprehensive atlas of the developing human immune system. Find out more 👇 https://t.co/N6KT4IFXav

Shared by EMBL-EBI at 5/13/2022

Objective often matters more than architecture. UL2 uses a mixture of denoisers objective, and allows one to switch between different downstream modes (finetuning, prompting) to get the best of both worlds.

Shared by Neil Houlsby at 5/13/2022

Check out our #NAACL2022 paper "Lifting the Curse of Multilinguality by Pre-training Modular Transformers" where we propose X-Mod, a modular MLM. 📜 https://t.co/97TxrQrEz4 ⌨️ https://t.co/mroChWDr3p w\ @NamanGoyal21 @VictoriaLinML @xl_nlp JamesCross @riedelcastro @artetxem

Shared by Jonas Pfeiffer at 5/13/2022

Bayes rules is an excellent introduction to Bayesian statistics (and easier than BDA3), and this book club is a great opportunity!

Shared by Aki Vehtari at 5/13/2022

Cool to see startups come out of work done at Uber AI labs! Congrats and good luck Piero!!

Shared by Jeff Clune at 5/13/2022

GTA-Human HRM baseline model is officially released in #MMHuman3D !

Shared by Ziwei Liu at 5/13/2022

My kid's book testing the limits of visual recognition. How on earth can you recognize this is a fish if not from context!? Makes me think the aim to completely suppress the context influencing a prediction might not be the best goal to have.

Shared by Lucas Beyer at 5/13/2022

New preprint! This work introduces FETA, a benchmark for task transfer. FETA is the largest NLP benchmark for "intra-dataset" task transfer, where task transfer is isolated from domain shift. Paper: https://t.co/t23gzllIgC Code: https://t.co/LRiMSVo2n5 (1/7)

Shared by Alon Albalak at 5/13/2022

New work on designing a general multimodal transformer that generalizes across text, image, video, audio, time-series, sensors, tables, & set modalities, while improving the tradeoff between performance & efficiency: Paper: https://t.co/IEZVY6EowH Code: https://t.co/UGerfCXIG2

Shared by Russ Salakhutdinov at 5/13/2022

Lifting the Curse of Multilinguality by Pre-training Modular Transformers abs: https://t.co/rNzSbBrlmx github: https://t.co/BnZVgZnamp

Shared by AK at 5/13/2022

Our most general agent yet!! Fantastic work from the team!

Shared by Demis Hassabis at 5/12/2022

If you can tokenize it, you can train a large language model for it. “Tokens in -> tokens out” is a very flexible computing paradigm. What *can’t* be represented in this way?

Shared by Russell Kaplan at 5/12/2022

Terrific lecture from @brwilder for VLDD Best PhD Thesis Award @aamas2022 "AI for Population Health: Melding Data and Algorithms on Networks" ! Congrats! #AIforSocialImpact #AIforPublicHealth

Shared by Milind Tambe at 5/12/2022

🚀Deploying advanced ML technology to serve customers and business needs requires a rigorous approach and production-ready systems. See why MLOps is important for any product, and how TFX helps make it real with @robert_crowe at #GoogleIO. Watch now → https://t.co/7kMSc6i8pN

Shared by TensorFlow at 5/12/2022

Looking fwd to hosting @brwilder for his Victor Lesser Dissertation Award lecture @aamas2022. Bryan's thesis was on "AI for Population Health: Melding Data and Algorithms on Networks." The lecture starts in 20 min. https://t.co/F7ThGs9Fki @MilindTambe_AI #AAMAS2022 #AAMAS

Shared by Haris Aziz at 5/12/2022

Mini-🧵 on OOD detection. I think it was @ChrisBishopMSFT who first proposed to threshold the loglik of a ML model f(x) to classify x as ID/OOD. However, since @eric_nalisnick et al. seminal paper "Do deep generative models know what they don’t know?" https://t.co/byCd1XwDG7, we

Shared by andrea panizza at 5/12/2022

Openness in action. We released the largest ever open language model, within a week @huggingface has it up and available for anyone to use! Nice work @MetaAI and @huggingface !

Shared by Mike Schroepfer at 5/12/2022

Maybe I'm missing something? Despite being disappointed by the results, this is still great engineering and I'm excited for their future follow-ups. If you want to see an agent that achieves new goals zero-shot (although in the same env), check out LEXA https://t.co/M2Mjrb3BZM

Shared by Danijar Hafner at 5/12/2022

Meta AI is calling researchers and developers to submit their papers to be highlighted at our workshop at @CVPR. Our workshop will touch on explainable #AI for computer vision. Learn more: https://t.co/VVf7ZxyoWX

Shared by Meta AI at 5/12/2022

Very excited that @ApacheSpark won the SIGMOD System Award this year. Congrats to the whole community behind the project!

Shared by Matei Zaharia at 5/12/2022

Last week, I submitted a comment to the California State Board of Education, objecting to the proposed revision to California Math Framework: https://t.co/srlJVmwRoJ Still time to submit you own (deadline 5/16) Some tips on how to submit & my reasons for submitting: 🧵👇 https://t.co/DKTz6l0d6v

Shared by Yisong Yue at 5/12/2022

Neat work from DeepMind, training a generalist agent that can both control a robotic arm and output text results (e.g. for image captioning or Q&A) https://t.co/4C50NPOpCp

Shared by Ben Hamner at 5/12/2022

Martin Volk and @sinaj_gz from our instute @UZH_ch about the phenomenon called "algospeak" in @SRF Wissen: https://t.co/FSV5pFhACM #NLP #NLProc #Algorithms #srfwissen

Shared by Zurich Computational Linguistics Group at 5/12/2022

Fun fact: This boat is correctly classified by DNNs until you remove just *9* training points. We explore (https://t.co/0PgzFhnJYj) how to identify such *data-level brittleness* via datamodels (https://t.co/ypR7MTCrHS) w/ @andrew_ilyas @smsampark @logan_engstrom @gpoleclerc (1/2)

Shared by Aleksander Madry at 5/12/2022

Here's a look back at how AI research at @Stanford has evolved from scholarship focused on pursuing generality to a multidisciplinary field – drawing experts in medicine, psychology, economics, law, art, and humanities. #ThrowbackThursday https://t.co/YdsUEoJKfc

Shared by Stanford HAI at 5/12/2022

📄 Check out how @Mitch_P and @RolandDunbrack visualized their research with Rascore, an app for analyzing the 3D structural models of RAS proteins. 🔬 📖 Read more: https://t.co/w6v9ZSRt2S #python #biology #cancerresearch

Shared by streamlit at 5/12/2022

New work shows that active offline policy selection (A-OPS) can accelerate policy development in real-world applications like robotics. A-OPS helps to quickly identify the best policy even when evaluation time on the robot is very limited: https://t.co/slBMBiIc5U 1/

Shared by DeepMind at 5/12/2022

The world is coming to Edmonton for Amii's inaugural AI Week, May 24-27th. The latest issue of @DisruptionMagCA has a sneak peek at what you can expect, as well as profiles of two of this year's keynote speakers: https://t.co/6z59izaoah #AI #ML #ArtificialIntelligence

Shared by Amii at 5/12/2022

We are happy to announce that @BonPress will sponsor our best paper award. More excitingly, accepted papers at #AdvML2022 workshop now have the option to publish in the Journal of Computational and Cognitive Engineering <https://t.co/csAqrKZ9WJ> with a fast-track review process!

Shared by Pin-Yu Chen at 5/12/2022

Last call for submitting your doctoral ideas at Hybrid #iswc_conf #iswc2022 Doctoral Consortium Tracks! A great opportunity to receive constructive feedback from mentors and most importantly your peers in the field! Click here for more information👉https://t.co/tVhI2irbu1

Shared by International Semantic Web Conference at 5/12/2022

Congrats to the team at @MGnifyDB on their 2.4B sequence protein database release! We're glad to have supported this effort with an #ML generated collection of 1.5B protein function predictions. Read more about this project, including prior work, below: https://t.co/5BtFo1iXAo

Shared by Google AI at 5/12/2022

We are very delighted to announce that our #graph4nlp long survey paper (190 pages) has been accepted by the Foundations and Trends® in Machine Learning Journal (https://t.co/2JOrndnqh4). Pre-publication link: https://t.co/axROWuFrne #GNNUpdates #GNN #NLP #NLProc

Shared by Lingfei Wu at 5/12/2022

I have been testing this branch with scikit-learn and there are still some open issues, e.g.: https://t.co/r8RkiORgiL but @colesbury is very fast at fixing them, sometimes in third-party libraries like numpy and Cython and this work will hopefully be upstreamed.

Shared by Olivier Grisel at 5/12/2022

Running 30B models in colab was certainly a challenge, but OPT models by @MetaAI from @stephenroller and team were a great test case for the feature!

Shared by Lysandre at 5/12/2022

So excited to share what I've been working on this past month. Not enough RAM to load the model? No problem. Not enough GPU RAM to host the model? No problem. As long as it fits on your hard drive, we'll make it run.

Shared by Sylvain Gugger at 5/12/2022

@MetaAI's OPT models come in flavors: from 125M to 175B params. Models up to the 30B variant are freely accessible, Accelerate v0.8 breaks the 6B parameter limit on colab, enabling: - Up to 11B in free Colab - Up to 30B in Colab pro Model card: https://t.co/0T0DYoleNn

Shared by Hugging Face at 5/12/2022

▶️ 𝚃-f𝚎𝚠 + (𝙸𝙰)³ ◀️ ---> "Few-Shot Parameter-Efficient Fine-Tuning is Better and Cheaper than In-Context Learning"! 🧵👇 https://t.co/zXSouyMdro h

Shared by Mohit Bansal at 5/12/2022

Months ago, OpenAI's Ilya Sutskever tweeted that today's neural networks may be slightly conscious. In our latest podcast episode, @spaniel_bashir talks to celebrated philosopher David Chalmers about consciousness and AI systems. https://t.co/QNrFXQJptm

Shared by The Gradient at 5/12/2022

Model FLOPs utilization (MFU) introduced in Pathways LM work (https://t.co/5K8SIwk9WD, sec 4.1). MFU is roughly the ratio of the measured iterations-per-second to the theoretical max (for a given hardware). Mostly designed to penalize activation caching / redundant work.

Shared by Awni Hannun at 5/12/2022

Learn how Intel #AI Builder @Gi_Bots is automatically processing invoices with a #DeepLearning solution. Their real-time processing engine uses Intel #XeonScalable processors to achieve low latency analysis of images and text. https://t.co/ACnZ9Ch3Zb

Shared by Intel AI at 5/12/2022

Another stunning BH image! I am surprised by the orientation of the galactic center black hole. Has anyone modeled potential accretion histories and investigated the likelihood of BH spin and galaxy spin?

Shared by David Spergel at 5/12/2022

Self-supervised trajectory (sequence) modeling learns generalist agents that succeed at multiple tasks! 🔥 Another feather in the cap for task-agnostic pre-training with sequence models (DT, LLMs). 🚀

Shared by Aravind Rajeswaran at 5/12/2022

I like the TPU vs GPU competition, and MFU as a metric is pretty healthy. Congrats to Google and NVIDIA for hitting 50%+ MFUs on their large training runs: https://t.co/P5ywgxhP3o https://t.co/tOepgknGUV

Shared by Soumith Chintala at 5/12/2022

New UL2 model/paper from @GoogleAI! "Unifying Language Learning Paradigms" ✅ SOTA on 50-ish diverse NLP tasks ✅ Outperforms GPT-3 175B on 0-shot SGLUE ✅ 3 x perf vs T5 XXL (+LM) on 1-Shot XSUM ✅ Open code & 20B Flax checkpoints. Paper: https://t.co/U1Ks1uNUCz

Shared by Yi Tay at 5/12/2022

🐈 Our most general agent yet? The exact same network and weights can generate text caption images, play Atari, control a robot, and more! Excited to be able to share this and see what comes next 🎉

Shared by Gabriel Barth-Maron at 5/12/2022

Honestly, it’s scary that data folks don’t know why they should model data. Data modeling provides the meaning/knowledge/semantics of the data. Without it, we don’t know what the data means. Excellent thread by @ergestx on why you need data modeling. Excited about his course

Shared by Juan Sequeda at 5/12/2022

We're always looking for more general agents, but this one takes the cake - it can play Atari, stack blocks with a real robot arm, and caption images. With the same model. And the same outputs. At the same time. Awesome! 🐈‍⬛😻 Congrats @scott_e_reed and @NandoDF and team! h

Shared by raia hadsell at 5/12/2022

Excited to receive an Early Researcher Award from Ontario (Canada) in support of the CleverHans lab's research on trustworthy ML! Very thankful for all the contributions of my group members, this award is a recognition of their creativity and hard work :) https://t.co/2azCEaycRr

Shared by Nicolas Papernot at 5/12/2022

This release represents a monumental amount of work for each one of these languages, and we have 67 new ones (plus new annotations/tasks)! 🎉🎉

Shared by Antonis Anastasopoulos at 5/12/2022

How does the UK currently translate AI R&D into goods, services, and businesses? What policies can improve this process? We have just published a report with @OfficeforAI and @CambridgeEcon on the UK's AI commercialisation process. Highlights below 🧵👇 https://t.co/XgEm7jhwNj

Shared by Oxford Insights at 5/12/2022

#NLProc hivemind: Does anyone know papers that discuss or study the trade-offs involved in deciding on how many tokens to include in your model vocabulary? I'm looking for hints about why the researchers settled on 30,000 (BERT), 50,257 (GPT-2), or 250K tokens (XLM-R)?

Shared by Desmond Elliott at 5/12/2022

My mind is blown by new UL2 model. - UL2 20B beats GPT-3 175B on zero-shot SuperGLUE - Public 20B model - Also finetunes really well paper: https://t.co/7HNMUex99s

Shared by Jason Wei at 5/12/2022

"Synergy of Prediction and Control in Model-based Reinforcement Learning" My thesis is now available, thank you @Berkeley_EECS and all who contributed to this. It has been a massive journey. https://t.co/wFfEucnHw0 I'll explain what I wrote about and a few reflections 👇 1/

Shared by Nathan Lambert at 5/12/2022

Few-shot in-context learning induces a language model to perform a task by feeding in a few labeled examples along with a task description and then generating a prediction for an unlabeled example. Processing these “in-context” examples can incur a huge computational cost. (2/9)

Shared by Colin Raffel at 5/11/2022

Hanging out with @KostasPenn! I’ve been very fortunate to have such AWESOME mentors and role models. Future home of the @Penn #datascience building.

Shared by Kosta Derpanis at 5/11/2022

Super excited about such crosslingual capabilities of the PaLM model (without any explicit crosslingual supervision). Here are some examples in Bengali, my native language.

Shared by Dipanjan Das at 5/11/2022

Please consider submitting your work to @cikm2022 , the abstract deadline for full and applied research papers is May 12!!! #cikm2022 #NLProc #SIGIR

Shared by julia kiseleva at 5/11/2022

@krasul and @NielsRogge are 100% the best people on time-series (w or without transformers)

Shared by Julien Chaumond at 5/11/2022

I am looking for a postdoc to work on computer vision & data science for monitoring moths & other insects. Position at @Mila_Quebec in collaboration with entomologists around the world. Should have PhD in CS, stats, ecology, or other related field. Email me if interested.

Shared by David Rolnick at 5/11/2022

Nice to see an empirical study on what humans can actually do in terms of one-shot generative generalization: One-shot generalization in humans revealed through a drawing task https://t.co/m7pf9BzvdK

Shared by Blake Richards at 5/11/2022

Forbes has identified @anacondainc CEO @pwang's PyScript @pyscript_dev announcement as a standout moment from PyCon US! #python #pyscript #datascience https://t.co/0OpipdulXK

Shared by Sophia Yang at 5/11/2022

🗣️💬🤖CLUES is a new benchmark for classifier learning using natural language explanations. It consists of a range of classification tasks over structured data along with natural language supervision in the form of explanations. https://t.co/lGUUi2C8Pn

Shared by Papers with Datasets at 5/11/2022

Say hello to the future of NLP topic modeling! 🤖 NLP transformer embedding 🧠 Advanced dim reduction + clustering with UMAP and HDBSCAN 🔖 Identifies topics using c-TF-IDF @MaartenGr's BERTopic does all this + more in a few lines of code 👏 https://t.co/Dk6SuTGh28 #NLProc

Shared by James Briggs at 5/11/2022

In this tutorial, Jakob Nissen gives non-professional programmers a brief overview of the features of modern hardware that must be understood to write fast #JuliaLang code. https://t.co/6Buc7rMhmP

Shared by The Julia Language at 5/11/2022

📣 The team of @nebulastream just released version 0.2.0. The system aims to unify data management approaches that are realized in different systems: cloud-based streaming systems,fog-based data processing systems & sensor networks. 👉Read more: https://t.co/1A8z5o9aDT #DIMA #AIM

Shared by BIFOLD at 5/11/2022

🌟🎉Special congratulations to Prof. @radamihalcea, her 2006 paper “Corpus-based and Knowledge-based Measures of Text Semantic Similarity” was recognized by @RealAAAI w/ a Classic Paper Award Honorable Mention for lasting impact on #NLP: https://t.co/UbsqzECVYN

Shared by MichiganAI at 5/11/2022

Nice collection of the ways reinforcement learning is being used to solve real-world problems… it’s getting better and better, finally! https://t.co/RfZilJKTm1

Shared by Mustafa Suleyman at 5/11/2022

Replay Grief is a story about re-animating the dead using generative models. It reads like scifi, but it's actually recent history - e.g, 'I gave my microwave a soul' https://t.co/FYJHjUZVpK and 'The Jessica Simulation: Love and loss in the age of A.I' https://t.co/zaZ4dW1SiI https://t.co/vDgNW5Lbyi

Shared by Jack Clark at 5/11/2022

Bursty data + transformer -> in-context (few-shot) learning not weight-based learning Bursty data + LSTM -> weight-based but no few-shot learning Zipfian (long-tailed) bursty data + transformer -> both in-context learning AND weight-based learning https://t.co/MPjEYXQ1Y5

Shared by Felix Hill at 5/11/2022

A bit over a week left to apply for PhD positions in #ML #IR #NLProc @DIKU_Institut for a start in autumn 2022. PhD topics include: interpretability, overparameterisation & generalisation, Web & Information Retrieval Apply here by 19 May: https://t.co/pJiUQ0ffU5

Shared by Isabelle Augenstein at 5/11/2022

Exciting webinar tomorrow by our very own Wilker Aziz and special guest André Martins on "sparsity in deep learning".

Shared by AmsterdamNLP at 5/11/2022

Nice to see @andrewyates talking at #voginip - bringing his research on neural IR to the Dutch community

Shared by Paul Groth at 5/11/2022

2): Replacing layernorms with batchnorms to avoid the inference-time overhead of the former. A one-for-one replacement makes the training diverge, so they had to add in more batchnorms and do so in particular places. [4/6]

Shared by Davis Blalock at 5/11/2022

🎉Ep. 5 of Neural IR Talks is out! Conversational Search w/ special guest @tonylittlewine and cohosts @andrewyates @SergiCastellaSa ❓We discuss Conversational Search focusing on the ConvDR paper from Shi Yu et al. from @MSFTResearch & @Tsinghua_Uni https://t.co/aRNg8JplYt

Shared by Zeta Alpha at 5/11/2022

You cannot take the logarithm of the Lipschitz constant of a function! A 🧵 about a super common mistake in ML papers 1/10

Shared by Francesco Orabona at 5/10/2022

⚡️ PyTorch Lightning 1.6 is the work of 99 contributors who have worked on features, bug-fixes, and documentation for a total of over 750 commits since 1.5. 🎉 This is our most active release yet! 👉 Here are some highlights: https://t.co/3HNDmBvbUy

Shared by PyTorch Lightning at 5/10/2022

A new Consensus Report from the National Academies @theNASEM: Automated Research Workflows Are Speeding Pace of Scientific Discovery. It has a few pages devoted to workflows powering simulation-based inference & reinterpretation (RECAST) in physics https://t.co/ui3nmSFHhQ

Shared by Kyle Cranmer at 5/10/2022

📈 Join the "How and Where to Publish Your Data Online" virtual workshop by @HMSCountway on May 11 at 12:00 PM EST for a hands-on experience preparing #data for publishing by “curating” an example dataset and identifying common data issues. Register now: https://t.co/pJjOASplrP

Shared by Harvard Data Science Initiative at 5/10/2022

The huggingface_hub library is the most underrated @huggingface project. It's the easiest way to share your ML models with your team or the community. Give it a⭐️: https://t.co/oiOjdJHM3c Here's an example from the wild - sharing a model from LucidRain's vit_pytorch package🚀

Shared by Nate Raw at 5/10/2022

Took some examples from the "ImageNet-trained CNNs are biased towards texture" paper. Nice to see @DeepMind Flamingo 🦩🦩 showing the desired shape bias! 😍 (without explicitly being trained to do so!) Glad to be a part of this! We've come a long way with deep learning 🧵👇 1/

Shared by Aleksa Gordić at 5/10/2022

Amazing venue to discuss the latest advancements in population models for large-scale #neural recordings! #neuroscience #ML Thanks Nina and Arno! Details here👇👇👇

Shared by ANC@Edinburgh at 5/10/2022

Easy: according to certain people (not me), a symbol is whatever deep learning currently cannot process but should 😂

Shared by Christian Wolf at 5/10/2022

Join @BostonNLP on May 23 for Optimizing Transformer Models for Performance, presented by Mark Kurtz from @NeuralMagic, hosted by @Kensho in Cambridge MA. We'll also support remote, Zoom participation. https://t.co/a5cnebqWam #NLPproc #meetup #ML

Shared by Seth Grimes at 5/10/2022

Super looking forward to this! Coming up this Thursday on the MMLL webinar series: Prof. @andre_t_martins will tell us all about sparsity in DL (from representations to distributions to communication channels).

Shared by Wilker Aziz at 5/10/2022

We're looking forward to seeing everyone 3 weeks from now in Crete. We'll be discussing the latest in #semantictech #knowledgegraphs & #webdata research. Connecting with colleagues in-person. https://t.co/zrucuGhEh8

Shared by ESWC Conferences at 5/10/2022

Optimum v1.2 adds ACCELERATED inference pipelines - including text generation - for @onnxruntime🚀 Learn how to accelerate RoBERTa for Question-Answering including quantization and optimization with 🤗Optimum in our blog 🦾🔥 📕https://t.co/CMSdeV0bLd ⭐️https://t.co/GwAHZMWa0a

Shared by Philipp Schmid at 5/10/2022

A quantum hardware startup and a quantum software startup are joining forces through an acquisition.

Shared by Tom Wong at 5/10/2022

It's my first post-Covid travel & I'm going to Princeton ☀️ In my "New Languages for NLP" keynote, I'll talk about developing advanced NLP pipelines for diverse languages, @spacy_io's philosophy & more! 📆 May 12, 4:30pm ET / 22:30 CEST (live stream) https://t.co/0VoPc5q1JZ

Shared by Ines Montani 〰️ at 5/10/2022

Awesome @QuantaMagazine @walkingthedot article on interpretable machine learning for science https://t.co/xSCC45wffk Guest appearances by @rogertgn @laurezanna @hodlipson, myself, and others. Discusses methods such as SINDy (@eigensteve++), PySR, Eureqa, NN interpretation w/ SR!

Shared by Miles Cranmer at 5/10/2022

Finally, we released the code, verbalizations and pre-trained models in order to easily reproduce the results! Models: https://t.co/yYq2chDRkB Thanks to @huggingface for hosting the models!

Shared by Oscar Sainz at 5/10/2022

What does it take to reproduce not only behavior but also large-scale electrophysiology across 9 different labs? Find out in this preprint! standardization procedures, quality control criteria, data, and analysis code all available. Congrats to the @IntlBrainLab e-phys team! 🧵

Shared by Surya Ganguli at 5/10/2022

Our #CVPR22 paper on Panoptic NeRF is now released on arxiv. TL;DR: view synthesis + semantics on “stuff” and objects in the scene. Object-based NeRFs also allow for editing/moving/removing.

Shared by Frank Dellaert at 5/10/2022

Progress in AI has been, and continues to be, extraordinary! But there's a still a long way to go to match human intelligence. We should hold both those facts in our minds 9/9

Shared by Murray Shanahan at 5/10/2022

Great to welcome Nouamane in my research team 🤗. We work together with Loic on a project we name "Massive" - which will be the most comprehensible study of text embedding models 🔢. Stay tuned for great charts📈, many insights 🔍 and a new understanding of embedding models🤯.

Shared by Nils Reimers at 5/10/2022

“A Tutorial on Structural Optimization” Blog: https://t.co/wagVSd76Up We implement structural optimization in 180 lines of Python. Then we use it to design trusses, bridges, and buildings.

Shared by Sam Greydanus at 5/10/2022

We have received a verdict from a jury in the Circuit Court for Fairfax County, Virginia, awarding Appian $2.036 billion in damages from Pegasystems Inc–the largest damages award in Virginia state court history. https://t.co/2g06Dm0LlW

Shared by Appian at 5/10/2022

Are your employees satisfied with the onboarding process? With #RPA optimizing HR workflows, one global bank could onboard new employees 70% faster, saving over $1M in the process. 📺 https://t.co/q1sfag553U

Shared by Automation Anywhere: The only cloud-native #RPA at 5/10/2022

I've joined @MSFTResearch Amsterdam as a Senior Researcher! As part of an amazing new team lead by @vdbergrianne and @wellingmax I will be working on reinforcement learning and deep learning for molecular simulation. Looking forward to working in this space!

Shared by Elise van der Pol at 5/10/2022

The way the community has responded to this has been beyond incredible. @ariG23498 and I are still processing the depth of the support we received in all honesty. Here's an update. 1/

Shared by Sayak Paul at 5/10/2022

Apply here by May 11th https://t.co/3aJkNW1uus. Call for NAACL scholarship applications to the undergraduate summer school at the 2022 Annual Jelinek Memorial Workshop On Speech And Language Technology (JSALT)! Please help spread the word! #NLProc

Shared by Diyi Yang at 5/9/2022

In the latest blog post, Mila's @Vincent_Mai_, @hbutsuak and professor Liam Paull @duckietown_coo present their paper, "Sample Efficient Deep Reinforcement Learning Via Uncertainty Estimation" which was presented at #ICLR2022 in the spotlight track. https://t.co/WfShthQMro

Shared by MilaQuebec at 5/9/2022

How can machine learning be used to leverage limited data from present-day quantum computers? We demonstrate the enhancement of variational Monte Carlo simulations using typical data from Rydberg atom arrays @Perimeter @UWaterloo @Stanford https://t.co/CVTIisxszq

Shared by Roger Melko at 5/9/2022

🤔 Curious about the https://t.co/EIsJyyUwAA #ConversationalAI platform? Now’s your chance to explore the benefits https://t.co/EIsJyyUwAA can offer your business with a FREE trial. Get started today. https://t.co/2f0CJbG9q9 #CX #CustomerExperience #Chatbot

Shared by Kore.ai at 5/9/2022

So grateful to see @GiadaPistilli join the team! I've loved Giada's work on @BigscienceW and recent analysis of GPT-3 behaviors, super looking forward to seeing what comes next 🤗

Shared by Yacine Jernite at 5/9/2022

Congratulations to the newest cohort of Vector Scholarship in Artificial Intelligence (VSAI) recipients for 2022! VSAI recognizes top candidates pursuing AI-related master’s programs in Ontario. To read more click here: https://t.co/i1brmwWRP9 #scholarship #ai

Shared by Vector Institute at 5/9/2022

With @CrowdCent these past weeks we have been focusing on developing new robust @numerai Signals models. This yielded new components for Signals data loading and preprocessing, which we are open-sourcing in NumerBlox. Highlighting 7 of them below. https://t.co/MkKMRIvXkP (1/9)

Shared by Carlo Lepelaars at 5/9/2022

Series C! 🚀🔥🎉 Super proud of the team and the impact we are having. Pumped to continue making machine learning more transparent, open, and collaborative and continue building with the community! 🤗 And we're hiring for every position you can think of :)

Shared by Victor Sanh at 5/9/2022

Combing For Insight in 10,000 Hacker News Posts With Text Clustering https://t.co/y6ryQRwOIL New blog post! I embedded and clustered the top HN posts looking for insight on personal/career development. I built an interactive map and found ~700 posts that fit the bill. 1/n

Shared by Jay Alammar at 5/9/2022

Some extremely exciting news!🤗 We are raising $100 million looking forward to building the future of open Machine Learning, from Computer Vision to Reinforcement Learning. The future of ML is collaborative.🤗 Jobs: https://t.co/ZvxbxglWp5 Announcement: https://t.co/TnciFCprvH

Shared by Omar Sanseviero at 5/9/2022

Tomorrow we cover a very hot topic! @jo_brandstetter presents his #ICLR2022 spotlight paper "Message Passing Neural PDE Solvers"! https://t.co/OWp4YwkoGr Join us this Tuesday on Zoom at 3pm UTC: https://t.co/9fxz75bW1o Co-authors: @danielewworrall @wellingmax

Shared by Hannes Stärk at 5/9/2022

Exciting research featured in @univgroningen news magazine. Great to have you @steveabreu7 here at @cognigron working within the @PostDigital_ETN network.

Shared by Groningen Cognitive Systems and Materials (RuG) at 5/9/2022

First-order reasoning and planning together with fast learning from low-level data requires: "an architecturally significant weaving of neural and symbolic parts, neither of which would be sufficient to solve the problem" @Grady_Booch at least for now.

Shared by Artur d'Avila Garcez at 5/9/2022

Evaluating Natural Language Generation (#NLG) 💬systems is a challenging task. In our #ACL22 paper, we present our contribution to addressing this challenge: RoMe, a robust metric for evaluating NLG📈. More details at: https://t.co/MGNdcrBTSC #machinelearning #conversationalai

Shared by SDA Research at 5/9/2022

A really nice read with lots of different perspectives on how to improve #benchmarking in #AI. There is a lot of great work happening. And #NeurIPS is again running the datasets and benchmarking track. Deadline June 9th! https://t.co/em56zMO4s9

Shared by Joaquin Vanschoren at 5/9/2022

This week on @therobotbrains we had @adampbry, Co-Founder/CEO of @SkydioHQ, building the most advanced AI for drones! We covered role of real and synthetic data, deep learning, consumer experience, commercial use cases, the arc of autonomy for drones. https://t.co/uRke1FhdMm

Shared by Pieter Abbeel at 5/9/2022

🚨🚨NEW BLOGPOST I write about the new family of Expressive Graph Neural Networks and what expressiveness entails in the broader context. It's full of colourful diagrams and intuitive explanations to the math behind the concepts ✨⚡️ Check it out: https://t.co/MnHGw9Gc0w

Shared by Rishabh Anand at 5/8/2022

What I would like to see more of: untraditional path --> PhD/academic lab/industry lab/machine learning role/conference participation. PhD/Academic lab/industry lab --> open to collaborate beyond traditional academia to help support independent researchers/community research.

Shared by Sara Hooker at 5/8/2022

How should you protect your machine learning models and IP? https://t.co/qcPKcxhryJ

Shared by Pete Warden at 5/8/2022

Returning from a fantastic workshop at Simons/Berkeley on MARL. My talk about multi-agent control/RL now available at: https://t.co/X6WYrzGdSC

Shared by Elad Hazan at 5/7/2022

Merci beaucoup for visiting @Mila_Quebec, showing off the new hardware #stellaquadruped and answering all of our (@MontrealRobots) questions. https://t.co/dfAZETB62l

Shared by Glen Berseth at 5/6/2022

Submitting to #NeurIPS2022? Want to include a thoughtful reflection on the limitations of your work but not sure how to begin? We designed REAL ML, a set of guided activities to help ML researchers recognize, explore, and articulate limitations that arise in their research.

Shared by Jenn Wortman Vaughan at 5/6/2022

This year at the @RetailInnovate in booth 1533, Haptik and @Zendesk will be featuring Conversational AI solutions for eCommerce like you've probably never seen before. Claim your free Expo Hall pass, here 👉🏼https://t.co/PRvXhlkTgW #Chatbot #ecommerce #IVA

Shared by Haptik at 5/6/2022

Grammatical markers are implicitly aligned in pre-trained multilingual encoders by encoding the same grammatical functions through the same subset of neurons across languages. This may help explain the "unreasonable" effectiveness of zero-shot cross-lingual transfer.

Shared by Edoardo Ponti at 5/6/2022

The "power of scale for prompt tuning" paper suggests that prompting catches up to finetuning at about 10b parameters. I wonder how the memorization curves would look going from T5-base to T5-XXL

Shared by David Dohan at 5/6/2022

So instead of human computations being strictly compositional, and instead of strict compositionality being a path to better AI, “systematicity may be a graded competency afforded by environmental and educational factors” (https://t.co/BW3yf6clt5).

Shared by Andrew Lampinen at 5/6/2022

4 examples on how chatbots can generate better results within the banking industry. #botsforbanking #chatbots #AI https://t.co/arxhtQxYuD

Shared by ServisBOT at 5/6/2022

1. Memorization Memorization during fine-tuning is interesting and under-explored. Pretrained models memorize datasets of lexical relations much faster than factual ones, and both faster than random relations. In most cases, prompt tuning cannot memorize large datasets at all.

Shared by Jacob Andreas at 5/6/2022

Loading glass is a tough and delicate job. Let robots do it 🦾 Learn how Matsunami Glass increased production by 50% and improved working conditions for its employees: https://t.co/SUdDe6WJ1v #glass #manufacturing #automation

Shared by Universal Robots at 5/6/2022

Great article in @ScienceMagazine by @SilverJacket on benchmarking in AI, covering @DynabenchAI as one way forward. So much important work to do in machine learning evaluation!

Shared by Douwe Kiela at 5/6/2022

#ICLR22 Is it possible to create digital twins by just watching videos alone? Please check out our RISP framework that allows solving this challenging inverse problem by integrating deep representation learning with differentiable simulation and rendering!

Shared by Chuang Gan at 5/5/2022

This robot spins, this robot runs! Be it snow, be it gravel, be it bricks or a broken motor. We achieved a record speed of 3.9m/s on the Mini-Cheetah while maintaining robustness on varied terrains using reinforcement learning. Find out more: https://t.co/lHGBTmT1u7 #Robotics

Shared by Pulkit Agrawal at 5/5/2022

5 benchmarks that tell you it's time to audit and optimize your ecommerce product quiz. A THREAD 🧵👇

Shared by Octane AI at 5/5/2022

In this #QuantumAITalks video → https://t.co/Vh4FGMR740, Murphy Yuzhen Niu talks about our research on designing and benchmarking a new type of pulse sequence in a realistic simulation of surface code to mitigate crosstalk errors.

Shared by Quantum AI at 5/5/2022

I'm amused: @semanticscholar has wrongly attributed two recent under-review papers to me on a topic that I have not worked on before but that I am working on now. It's an excellent, mysterious prediction, Semantic Scholar, and I appreciate the nudge, but these things take time!

Shared by Christopher Potts at 5/5/2022

The AI Economist uses reinforcement learning to learn dynamic tax policies that optimize equality + productivity in simulated economies, outperforming alternative tax systems. Our expanded research has now been published in the journal Science Advances. https://t.co/5ZhE0xaYB7

Shared by Salesforce AI Research at 5/5/2022

✨NEW PAPER ALERT ✨ We propose EAR 👂 a new Entropy-based attention regularization term to prevent lexical overfitting in #Transformer models (Findings of ACL 22) Code: https://t.co/IAKx2r9xof By @peppeatta @debora_nozza @dirk_hovy @ElenaBaralis #acl2022nlp #NLProc

Shared by MilaNLP at 5/5/2022

The @MSFTQuantum team has created a new class of quantum error correction codes, called Floquet codes. Learn how the codes lower spacetime overhead by 10 times and substantially reduce requirements for a key process in the topological quantum stack: https://t.co/I2fSkyI1pL

Shared by Microsoft Research at 5/5/2022

If you think a private, cost-effective and accurate transcription microservice is impossible, try Leopard with gRPC! Customize Speech-to-Text models and transcribe 100 hours per month for free! It takes 3 mins to learn! https://t.co/zzsQixC1z8 #VoiceFirst #SpeechRecognition

Shared by Picovoice.ai at 5/5/2022

Nice things about EAR 👂: - it is applied to all tokens; no need for lists - it doesn't require model changes - code is available and transformers-ready Work w/ @debora_nozza, @dirk_hovy, and @ElenaBaralis Preprint: https://t.co/FYATayckoV #NLProc [4/4]

Shared by Giuseppe Attanasio at 5/5/2022

Day 4 of #ConstructionSafetyWeek: Continue Learning. A safe way of learning about the facility, its safety equipment and emergency exits, is a virtual record of the facility with hints, documents and more. See how 360° learning environment can look like: https://t.co/z1hWLczicc

Shared by HoloBuilder at 5/5/2022

We are proud of our JHU CLSP turnout for the 9th Mid-Atlantic Student Colloquium on Speech, Language and Learning (MASC-SLL 2022). Thank you @TempleUniv for hosting! #masc-sll See 👇 for some of our students’ work presented at the event.

Shared by JHU CLSP at 5/5/2022

Today (5/5) 4pm EDT, Ruslan Salakhutdinov @rsalakhu from CMU @SCSatCMU @mldcmu will be giving a talk at the Embodied Intelligence Seminar @MIT_CSAIL on building autonomous embodied agents! Live-streaming from: https://t.co/bQToLGI0qY…, co-hosted with @EpisodeYang!

Shared by Yilun Du at 5/5/2022

Amelia is hired, again! Hoffman Financial Group, a Georgia-based independent financial planning firm founded by Chris Hoffman, onboarded Amelia as a virtual assistant to handle all incoming prospect calls, capture lead information and more. #AmeliaHired https://t.co/o7wvrUbU5w

Shared by Amelia at 5/5/2022

Implementing SegFormer in PyTorch https://t.co/37ImTV8suO #deeplearning #machinelearning #ml #ai #neuralnetworks #datascience #PyTorch

Shared by PyTorch Best Practices at 5/5/2022

Tired of approximating the p-value of the paired-permutation test? Our new paper gives an 𝑒𝑥𝑎𝑐𝑡 algorithm for a common class of paired-permutation tests that is 𝑓𝑎𝑠𝑡𝑒𝑟 than Monte Carlo approximation! https://t.co/Px7lhUhHmF with @xtimv @ryandcotterell at @naaclmeeting

Shared by Ran Zmigrod at 5/5/2022

In this episode of @therobotbrains, @adampbry, Co-founder and CEO of @SkydioHQ, and @pabbeel discuss real-world applications for AI-powered drone technology. #artificialintelligence #drones #AI YouTube: https://t.co/ydnntHKi59

Shared by Covariant at 5/4/2022

For tomorrow's Stanford NLP Seminar, we're delighted to host Eunsol Choi (@eunsolc) from @UTCompSci speaking on Information-seeking Language Understanding in a Dynamic World. Join us over zoom tomorrow at 11 am PT. Reg: https://t.co/jnx3sHCfYG; Abs: https://t.co/ek4LHhLEbK

Shared by Stanford NLP Group at 5/4/2022

Even when benchmarking on OGB datasets, researchers have an incentive to maximize the evaluation metric, and not produce generalizable scientific insights. 4/9

Shared by Anton Tsitsulin at 5/4/2022

Large-scale character animation models: train a giant GAN on huge amounts of data to get skills, then train a small high-level controller to pick the skills to solve a task. Building on our prior adversarial motion prior work (AMP), this allows general-purpose animation models.

Shared by Sergey Levine at 5/4/2022

The Gu Lab @UCBerkeley @BerkeleyME has postdoc positions available in additive manufacturing, machine learning, and computational mechanics. More information here: https://t.co/0kTj6jq9si Please help RT, many thanks!

Shared by Grace Gu at 5/4/2022

We’ve been training giant neural networks that do stuff for you on your computer! In the first three months at Adept, we taught it to query databases, make visualizations, and fetch data from the web, but we want to teach it how to use every software tool in the world. https://t.co/2fkRVOgC9p

Shared by David Luan at 5/4/2022

In our first few months we (@AdeptAILabs) trained a neural network to do a bunch of different things on a computer, including querying databases, making plots, and fetching data from the internet. The video is in the quoted tweet, but here’s a few things that I thought were cool:

Shared by augustus odena at 5/4/2022

ICYMI - Here is a replay of my talk on #AI trends, state of the state, challenges of AI-enabled businesses and how to operationalize #artificialintelligence. Let me know what you think. https://t.co/YM6njQZpBQ #Verta #MachineLearning #ML

Shared by Manasi Vartak at 5/4/2022

Fail* 64x, Restart* 44x, hang* 28x, lost* 21x, kill* 15x ... Training large models is messy and hard... #OpenScience by @MetaAI https://t.co/XBLa9YSzwu

Shared by Jerome Pesenti at 5/4/2022

Nice and especially appreciate the release of the accompanying logbooks, detailing the struggles of training transformers at scale https://t.co/GErHySLdCJ

Shared by Andrej Karpathy at 5/4/2022

Today I'm hosting two Kaggle legends: @Tiwarihere and @Rob_Mulla We will learn about: - Rob's journey to becoming a Kaggle creator - Rajneesh's journey to one of India's top Kagglers - 3rd Pos solution to the music classification comp https://t.co/tgyCzDOhvs

Shared by Sanyam Bhutani at 5/4/2022

Looking back at the work we’ve done back in 2006 on measuring text semantic similarity, it would be fair to say “this is obvious” — and surprisingly it feels good to have done work that in retrospect appears as “obvious” ✨

Shared by Rada Mihalcea at 5/4/2022

Mila(https://t.co/cVXe4xYaEh) has multiple postdoc positions on ML for molecular modelling with applications in drug and material discovery. Candidates will be co-supervised by Mila professors Yoshua Bengio, me, Guy Wolf, ... Please DM me or email me (tangjianpku@gmail.com).

Shared by Jian Tang at 5/3/2022

@dlmcguinness @RobertMMetcalfe @BranaRakic @DrevZiga @cornell_tech #networks #graphs

Shared by Knowledge Graph Conference at 5/3/2022

Search apps for coding efficiency: 🏫 ArXiv 📦 AWS 💾 Code Complete 🎨 Color Picker 🤓 Geeks for Geeks 🐈‍⬛ GitHub 🤗 Hugging Face ✅ JSON formatter 🪟Microsoft Docs Ⓜ️ MDN 🧱 PyPi 🔥 PyTorch 📜 Read The Docs 📂 StackOverflow 👓 Towards Data Science 📗 Tutorials Point

Shared by Richard Socher at 5/3/2022

New jobs in the 21st century: Model restart specialist Hyperparameter psychic Prompt engineer Model janitor Tensor shape mediator Quantum state observer Model footprint accountant

Shared by Christian Szegedy at 5/3/2022

Still a couple of days to apply for a Senior Lecturer / Lecturer position (Asst Prof) in Machine Learning, in new Manchester Centre for AI Fundamentals, CS Dept. DL May 5.

Shared by Samuel Kaski at 5/3/2022

An applaudable effort from Meta AI (kudos to @LukeZettlemoyer and his team) on releasing the parameters and training logs of GPT-3 scale pre-trained LMs. The extensive evaluation and discussion on ethics and limitations are also impressive. #NLProc https://t.co/NtF78Sb85I

Shared by Yu Su at 5/3/2022

In our latest issue, a team of @Berkeley_EECS researchers present a framework for protecting individual #privacy without losing information when deploying a #MachineLearning system for reliable uncertainty quantification. @ml_angelopoulos #privacy Read: https://t.co/Ng1bTfRqXi

Shared by Harvard Data Science Review at 5/3/2022

🔥 YOLOS is now available in @huggingface Transformers! 🤩One of the most elegant and simple object detectors out there: YOLOS = a Vision Transformer (ViT) trained using DETR's clever loss function https://t.co/v2cWHoF1xG Disclaimer: YOLOS != YOLO series for object detection

Shared by Niels Rogge at 5/3/2022

Speech News📢 We just opened two new exciting positions for deep learning in speech🔥 👉 Research Engineer to pretrain large audio Models: https://t.co/05MXJfh5nK 👉 Machine Learning Engineer for Robust Speech Recognition: https://t.co/eNfOGM6J6Q

Shared by Patrick von Platen at 5/3/2022

Really eager for today's graph reading group! @simonbatzner is rethinking how to learn dynamics of 3D molecular and crystal structures beyond message passing! Join the zoom link at 11AM EST or 3PM UTC: https://t.co/PfvDmYBhUi

Shared by Dominique Beaini at 5/3/2022

1/8 Confused about the California Math Framework controversy? Don't feel like reading a 900 pages document? Brian Conrad (math prof at Stanford) did it for you at https://t.co/lvAFes6YG9 Some highlights (or lowlights) include:...

Shared by Boaz Barak at 5/2/2022

Is it possible to train an object detector from scratch without ImageNet pre-training? If so, what do you loose by doing this? Any caveats to be aware of?

Shared by Tom Goldstein at 5/2/2022

I'm almost scared to ask, but what is EA and what does it have to do with language modeling? (out-of-touch person on the east coast)

Shared by Sasha Rush at 5/2/2022

Refine and improve language model outputs with *language* feedback. Pretty neat! Does it mean we losing out in performance by not allowing models to "think" more with self-refining its outputs?

Shared by Elman Mansimov at 5/2/2022

And another present: two more of our papers to look out for: 6️⃣ Square One Bias in NLP: Towards a Multi-Dimensional Exploration of the Research Manifold ✍️ by @seb_ruder, @licwu, Anders Søgaard

Shared by CambridgeLTL at 5/2/2022

*Masked Siamese Networks for Label-Efficient Learning* by @mido_assran @mcaron31 @imisra_ @p_bojanowski @armandjoulin et al. Neat new self-supervised method that combines "patch masking" with a siamese architecture and DINO-like training. https://t.co/iFiZgVhJhv

Shared by Simone Scardapane at 5/2/2022

One week left to register to the @HBP_Education's symposium at @uni_copenhagen. Learn what the @EBRAINS_eu platform has to offer for neuro research. Including talks on computational neuroscience by @apdavison and neurorobotics by @SilviaTolu1 🤖 Info: https://t.co/Wd09De9eZX

Shared by Elias Najarro at 5/2/2022

The Real Virtual Humans group offers research fellowships for Ukrainian PhDs/PostDocs/researchers at risk in CV/ML as well as software engineering positions for BSc/MSc students.

Shared by Gerard Pons-Moll at 5/2/2022

Why is logistic regression the most common approach to classification at scale? When should you start your project with this method? And why is it such a good choice for ranking at scale? https://t.co/KIe6ijrhNc

Shared by Cassie Kozyrkov at 5/2/2022

The EPFL #NLProc lab is hiring a postdoc in the areas of multimodal language and vision systems and open domain QA. Apply here: https://t.co/thb7I5266e Preferred start date: Aug 1st

Shared by Antoine Bosselut at 5/2/2022

This excellent tutorial on implicit differentiation and implicit layers (neural ODES, differentiable optimization, etc.) really brings together the techniques, upsides, and downsides in a clear way: https://t.co/ZvwA3hgHv4

Shared by Eugene Vinitsky at 5/1/2022

O brave new world, that has such language models in it! https://t.co/37zlZ77mAX

Shared by Melanie Mitchell at 5/1/2022

🚨 Dataset Release 🚨 We are proud to release Universal Proposition Banks 2.0 to facilitate R&D on cross-lingual NLU. 🌎23 languages 🌲span + dependency based universal SRL annotations 🪙Gold data included 🔗 https://t.co/LZUF73pNqI Welcome your feedback and contributions🤩

Shared by Yunyao Li at 5/1/2022

A good experiment on generative models (DALL-E2) trying to count. They can count up to 3 reliably. 4 is usually correct but 5 and higher seems to have frequent errors.

Shared by Alex Dimakis at 4/30/2022

8 years of progress in generative modelling. What a time to be alive

Shared by Taco Cohen at 4/30/2022

Congratulations to Heyu for winning the Best Paper Finalist award at the GroundedML Workshop @iclr_conf! 🎉🎊 Very well deserved & all the more impressive given she is an undergrad! CNAP extends @andreeadeac22's XLVIN to continuous action spaces, and should be released soon 🦾🤖

Shared by Petar Veličković at 4/30/2022

Thank you Sepp🍻 In this work, we propose U-NO, a neural operator architecture that nicely exploits the problem structure of maps between function spaces, and finally allows for deeper and memory-efficient neural operator models.🔥 A work of @Ashiq_Rahman_s joint w @zross_ 1/n

Shared by Kamyar Azizzadenesheli at 4/29/2022

My research focus will be on geometric deep learning inspired by (but not limited to) problems in structural biology, drug discovery and others. If that sounds appealing to you, please reach out! I am looking for postdocs, PhD students and visiting researchers! 2/2

Shared by Octavian Ganea at 4/29/2022

We introduce the work that offers an integrated view of these two dimensions, and the integrated view naturally leads to simple yet strong empirical solutions with top performances on a collection of OOD benchmarks from us (COI noted) w. @dhuangtuite @yong_jae_lee

Shared by HaohanWang at 4/29/2022

The Office for AI and our colleagues across government will be at #ODISummit2022 along with our partners from industry and civil society. We'll be discussing the major issues of the day – from ethics, trust and business models, to emerging tech - and we want you to join us. 👇

Shared by Office for Artificial Intelligence at 4/29/2022

"Which Language Evolves Between Heterogeneous Agents?" at @EmeComWorkshop. Work led by @MarieRobots "Emergent comm. between agents with different time scopes and perception, trained with imagination-based learning is not predestined." https://t.co/hkwEu2zj5b

Shared by Kory Mathewson at 4/29/2022

Weekend listening #ICYMI! @johnmccrae on using technology to support minority languages, @artboxhill on how art & neuroscience can complement each other & Anne O'Leary on the @EcoShowboat! Lou's mic was iffy. We managed! #InsightCulture #InsightSpotlight https://t.co/LTJhKtCAlP

Shared by Insight Centre at 4/29/2022

It has been a pleasure to co-supervise @Ahmed_AI035 Master thesis! Check out his great work on Graph Anisotropic Diffusion to discover how you can achieve efficient and global anisotropic kernels for message passing. Could this research direction be the future of Graph ML?

Shared by Gabriele Corso at 4/29/2022

As part of DeepMind's mission to solve intelliegence, we halved the learning rate every 2,500 optimization steps.

Shared by Ferenc Huszár at 4/29/2022

Join us today for the @iclr_conf Workshop "Anchoring Machine Learning in Classical Algorithmic Theory" Amazing line-up of speakers w/ @StefanieJegelka, @PetarV_93, Aarti Singh, @prfsanjeevarora, @cypaquette, @69alodi, Gintare Karolin Starts at 8:45am EST https://t.co/UNgM7QlgBn

Shared by Xavier Bresson at 4/29/2022

Congratulations to Debanjan ( @deepinevil ) for successfully defending his PhD thesis today! Debanjan worked on integrating background knowledge into question answering & dialogue systems to increase the factual accuracy of chatbots while allowing very natural sounding dialogues.

Shared by Jens Lehmann at 4/29/2022

🦩 is pretty wild! turns out you can: 1. take a frozen pretrained LM 2. perform some model-surgery on it with cross-attention layers to ingest tokens from a visual encoder (only 15% extra parameters) 3. voila! you get a visual-LM that can do few shot learning

Shared by Ankesh Anand at 4/28/2022

Thanks for having us on! Was fun talking about model sparsity and how conditional computation can be a useful for the future of LLMs.

Shared by Barret Zoph at 4/28/2022

1/ I've been slow to post, but the group was fortunate enough to share three papers this week at #ICLR2022: Learning 3D Representations of Molecular Chirality with Invariance to Bond Rotations (@keiradams, @lucky_pattanaik) : https://t.co/HETwjZhwMj

Shared by Connor W. Coley at 4/28/2022

Excited about meta-gradients (MGs)? join our intern @jelennal_ tomorrow at @aloeworkshop to hear about a systematic study of MGs in non stationary RL environments. We study how adding a context to MGs can assist in adapting to changes and visualise what do these MGs learn.

Shared by Tom Zahavy at 4/28/2022

If you're at ICLR now, check Generalizable Policy Learning in the Physical World workshop tomorrow. We will present Prompts and Pre-Trained Language Models for Offline Reinforcement Learning and will be happy to share its current improvements. We have some 😉

Shared by Sergey Kolesnikov at 4/28/2022

How the brain works is an incredibly hard and crucial question! We're happy to share new work from the FAIR team at @MetaAI that shines a new light on how the brain processes language. Check it out

Shared by Antoine Bordes at 4/28/2022

@MosaicML will be here! Come by and visit the booth if you're around...we're hiring amazing software engineers (Python, Golang, C/C++, Kubernetes gurus) who want to want to build the future of machine learning infrastructure. Check us out! https://t.co/16RRa2kYRp

Shared by Naveen Rao at 4/28/2022

Analysis of ML and NLP publication statistics from 2021. https://t.co/1pDefoBhnn #machinelearning #NLProc

Shared by Marek Rei at 4/28/2022

Cool work from colleagues at @DeepMind and collaborators approximating turbulent dynamics with convnets.

Shared by Danilo J. Rezende at 4/28/2022

Happy to share our paper “A Causal Viewpoint on Motor-Imagery Brainwave Decoding” at #ICLR2022 (@iclr_conf) Workshop on Objects, Structure and Causality (OSC). Stop by our poster tomorrow 😄 w/ Y. Panagakis, D. Adamos, N. Laskaris and Stefanos Zafeiriou 🧵👇 (1/🧵)

Shared by Konstantinos Barmpas at 4/28/2022

When you think you see life here, it’s all in your eyes, just like when we anthropomorphise neural networks. All in our head Dont get me wrong, it’s beautiful how human put meaning in everything that surround them But it’s important to stay aware we build reality in our mind

Shared by Thomas Wolf at 4/28/2022

In two weeks I'll give a general-audience talk on generative models at the "house of culture" in Girona (in Catalan; with a focus on speech/music generation). For those interested in attending and talking face-to-face, there is a free registration link: https://t.co/0oWGAgdOPt

Shared by Joan Serrà at 4/28/2022

eigenvector of J’J gets amplified exponentially, which causes the gradient to also be amplified exponentially. Thus the objective function’s gradient map is not Lipschitz along the optimization trajectory, even though the network’s gradient map and the hessian are both Lipschitz

Shared by Jeremy Cohen at 4/28/2022

As a co-organizer of ICML workshop "Principles of Distribution Shift (PODS)", I invite you to submit your interesting findings, hear about peers' findings, and hear from our invited speakers w/@zacharylipton @ElanRosenfeld @saurabh_garg67 @ShibaniSan @jamiemorgenste1 @risteski_a

Shared by Hossein Mobahi at 4/27/2022

🚨LIVE NOW: Join us for a @TwitterSpaces conversation on weaponized conspiracy theories and misinformation featuring researchers & psychologists including @abbieasr, @Sander_vdLinden & our own @_BGoldberg, moderated by @cwarzel. https://t.co/tcgQk930ZR

Shared by Jigsaw at 4/27/2022

How models respond to distribution shift is the primary question in ML / CV. But often 'distribution shift' is left vaguely or implicitly defined. This paper provides one intuitive framework for defining distribution shifts and evaluates ways of making models robust to it.

Shared by Sagar Vaze at 4/27/2022

"Model downloads of XLS-R (300m,1B,2B) spiked at 3,000 downloads per day during event" https://t.co/oYKv1uI2kj Thank you @huggingface @PatrickPlaten 🤗 Speech technology for everyone in all languages will become a reality, not just for research, but also for concrete products.

Shared by Alexis Conneau at 4/27/2022

I am co-presenting 2 @iclr_conf spotlight papers at Session 11 tomorrow: - "Geometric and Physical Quantities Improve E(3) Equivariant Message Passing" together with @robdhess @ElisevanderPol @erikjbekkers - "Message Passing Neural PDE Solvers" together with @danielewworrall

Shared by Johannes Brandstetter at 4/27/2022

We can recall acquired skills and generalize them to perform new tasks, even w/o any supervision. Can multi-task LMs (e.g., T0) also benefit from retrieval for cross-task generalization? We show it’s promising and propose such a method, ReCross.🧵 [1/7] https://t.co/7urLFQjhKC

Shared by Yuchen Lin at 4/27/2022

Could Transformers become the gold standard for symbolic (and even non-symbolic) regression ? Check out our latest paper to find out !

Shared by Stéphane d'Ascoli at 4/27/2022

We publicly released 185GB digitized high-resolution histology slides of renal cell carcinoma and their subtypes according to the consensus opinion of expert pathologists at Dartmouth-Hitchcock. More information about this dataset and how to download: https://t.co/5urN6y7B0K

Shared by Saeed Hassanpour at 4/27/2022

Great to see @ApacheTVM reach 8k stars and approach 700 lifetime contributors with expertise in ML, compilers and computer architecture. Optimizing and compiling ML models for a broad set of {Model x HW} is best supported and made future-proof by a strong community! 🙏

Shared by Luis Ceze at 4/27/2022

Looking back at this, it's amazing how much has happened in machine learning applied to procedural content generation in just the few years since this paper was published. Still, non-learned methods based on search, solving, and optimization have many advantages.

Shared by Julian Togelius at 4/27/2022

We would be talking about our work at @iclr_conf, to be presented as oral today. https://t.co/mcFp9wVQ0V Integrating information across different functional elements via a learned bottleneck @Mila_Quebec

Shared by Anirudh Goyal at 4/27/2022

Amazing text to speech project with generative machine learning model, check out the demos https://t.co/sbmKbNryIl

Shared by jtoy at 4/27/2022

We took home the Gold thanks to the team at @EdisonAwards. SensaAML won in the Commercial Technology, Data Security Solutions category. So extremely proud of our team for this amazing recognition. #innovator #teamwork https://t.co/C9FwGLV7Rw

Shared by Symphony AyasdiAI at 4/27/2022

Check out our Spotlight at #ICLR2022 -- @GeigerFranziska shows that 3 complementary techniques can reduce the number of supervised synaptic updates to train models of primate vision by *two orders of magnitude* while retaining 80% of the brain predictivity https://t.co/swfU7RH3Lq

Shared by Martin Schrimpf at 4/27/2022

This paper surfaces a very important problem in deep RL! 👏🏻 Really looking forward to reading it in detail.

Shared by Arash Tavakoli at 4/27/2022

Today at #ICLR22 : Deep RL agents have to fit a series of value functions -- we show that this can make neural networks **worse** at fitting new targets later in training, and explore the implications of this in deep RL. 📜https://t.co/uXHwayYD6r 📺https://t.co/BakmfWSniY

Shared by Clare Lyle at 4/27/2022

🚩Don’t panic when hearing this: there is high train-test overlap in MS MARCO and Deep Learning Tracks. Test queries are too similar to the training ones! How to resolve such bias? Do existing findings hold without the bias? Here is our recent work (1/9) https://t.co/rTTVBX8evV

Shared by Jingtao Zhan at 4/27/2022

We’ll be presenting WILDS v2.0 as an oral at ICLR! We extended the WILDS benchmark of real-world shifts by adding unlabeled data, which can be used for domain adaptation and representation learning. Talk + poster: https://t.co/VhCyKkHw9P Paper: https://t.co/Z47ek3YRcy 🧵

Shared by Shiori Sagawa at 4/26/2022

Nothing warms my heart more than seeing an exodus of #AI research folks from big tech,going into various startups. I truly believe that startups are our only way of making meaningful disruption of various industries, through AI systems that actually work in the messy real world!

Shared by Nasrin Mostafazadeh at 4/26/2022

Inspired by how biological neural networks are grown through a self-organized developmental process, we introduce: HyperNCAs 🧬🧠 - Growing Developmental Networks with Neural Cellular Automata PDF: https://t.co/Jdaa0hxRse Led by @EIiasNajarro w/ @SudhakaranShyam & @claire__aoi

Shared by Sebastian Risi at 4/26/2022

As an end user, you may never have to work with artificial intelligence from a developer’s perspective, but knowing how it works is key to ensuring its successful use. Fortunately, the technology can explain itself. https://t.co/VykHo6zSpS by @automationworld

Shared by Neurala at 4/26/2022

Really proud how widely the GSM8K Math Problems dataset (https://t.co/5MdDzjJ5SW) has been used in Google's PaLM and Chain of Thought papers! These math and Python datasets have been super interesting to create @HelloSurgeAI — moving far beyond basic "sentiment labeling".

Shared by Edwin Chen at 4/26/2022

Introducing SignNet and BasisNet: neural networks designed to handle the symmetries of eigenvectors.

Shared by Josh Robinson at 4/26/2022

5 time-series concepts you should know about: 1. ARIMA 2. Prophet @MetaAI 3. AutoML for Time-Series Forecasting @GoogleAI 4. DeepAR @AmazonScience 5. Temporal Fusion Transformers @GoogleAI Learn more about it here: https://t.co/lFm1TF8Phu

Shared by TheSequence at 4/26/2022

A new release for Quantus is out! 🎉 🎉 Among other things, this update includes new evaluation metrics: (In)fidelity and ROAD plus we are now using a more generous LGPL license model. https://t.co/wtYqS7WNNs

Shared by Understandable Machine Intelligence Lab at 4/26/2022

Great opportunity! I'll mentor a project together with @_onionesque on equivariant poset representations. If you're a bit tired of graphs and sets, come contribute to a new and promising combinatorial world :-)

Shared by Leonardo Cotta at 4/26/2022

Our PyTorch extension escnn for isometry equivariant CNNs is now upgraded to include volumetric steerable convolution kernels. https://t.co/iMn5FqMFPt Come to our #ICLR poster session today 👇

Shared by Maurice Weiler at 4/26/2022

31st Annual Computational Neuroscience Meeting (CNS*2022) Melbourne 16-20 July 2022 https://t.co/DQR4tjA0Td The CNS*2022 abstract submission deadline for posters has been extended to Monday, 02 May at 14:00 European Summer Time. Please note this is for posters only.

Shared by OCNS at 4/26/2022

We've also made the same scaling plot for MLP-Mixers to see what the effect of architecture was (@balajiln @jessierenjie) [https://t.co/uT0q1GZ5ZI]:

Shared by Stanislav Fort (back in SF) at 4/26/2022

In my view, logical reasoning is risky, because it detaches the conclusions from the context. Hence, it needs to be checked against the statistics of real-world experience.

Shared by Thomas G. Dietterich at 4/26/2022

The project co-led by my very first research mentee @yotaros_ramen and the awesome Alex Paunov is now out as a preprint! We find synchronized language network activity across people as they watch movie clips / audio event sequences with no one speaking in them. 1/n

Shared by Anna Ivanova at 4/25/2022

Let a bot be a bot. Let it acknowledge what it doesn't know but still be helpful and creative. Current dialogue datasets encourage the opposite and license models to hallucinate alarmingly (#naacl). FaithDial (preprint) is all about trustworthy chatbots https://t.co/xRyJWlFOWL

Shared by Siva Reddy at 4/25/2022

Excited to share a new blogpost I wrote about GPT-3 Prompt Engineering on @weights_biases Fully Connected! 🤖📕 Learn how to get GPT-3 to summarize text, do sentiment analysis or even sanity-check your ideas by designing prompts in plain English.🤯 Link: https://t.co/9MeGXIjbxI

Shared by Ivan Goncharov at 4/25/2022

Explainability is really important for AI models, but many existing techniques are not very useful, and improving them isn't easy. We propose a method to *learn* explainers by finding ones that better *teach* a student model. Really excited about this, check it out!

Shared by Graham Neubig at 4/25/2022

The @BigscienceW Architecture group found that ALiBi improves performance on a set of 27 downstream tasks! I will present the ALiBi poster tomorrow at @iclr_conf at: 6pm London Time 1pm Eastern 10am Pacific Come chat about making transformers better! https://t.co/caGpITBacT

Shared by Ofir Press at 4/25/2022

What it lets you do is tune most of your hyperparameters on a *much* smaller model than the one you care about. E.g., you can do your full BERT-Large hparam sweep with as much compute as a single BERT-Large training run. This is orders of magnitude less computation.

Shared by MosaicML at 4/25/2022

If you’re at #ICLR2022, hope you’ll check out our spotlighted poster: “Multitask Prompted Training Enables Zero-Shot Task Generalization.” https://t.co/kunkvJqsxu Poster session 5, Tue 1:30-3:30 ET

Shared by Stephen Bach at 4/25/2022

#ICLR2022 We present CP-Gen, a modular approach for improving the efficiency (e.g. length, volume) with conformal prediction, by tuning prediction sets with more than one parameters. Paper: https://t.co/Jg3gFsuMvY Poster (Monday 6:30pm PT): https://t.co/m7TrYHgEua

Shared by Yu Bai at 4/25/2022

Come see our #ICLR2022 paper on "Salient ImageNet: How to discover spurious features in Deep Learning?" on Mon 25 Apr 10:30 a.m. PDT — 12:30 p.m. PDT Link: https://t.co/Yzlo3HFda2 Dataset & code: https://t.co/nA8FcMmAIY joint work with @sahilsingla47

Shared by Soheil Feizi at 4/25/2022

I’m happy to announce that I have several open PhD positions in #deeplearning for #conversationalAI, #speech technologies, and more generally #MachineLearning for #sequence processing. More information here: https://t.co/7FenSZGv5S If you are interested, please apply!

Shared by Mirco Ravanelli at 4/25/2022

AI model drift is an important challenge in practice and is under-studied compared to data drift. Model updates can occur silently and disrupt the rest of the engineering pipeline for unclear reasons. Lots more to do! Great work led by Lingjiao Chen w/ @matei_zaharia!

Shared by James Zou at 4/25/2022

For those interested in NLP on African languages, check out our latest AfriBERTa follow-up examining vocabulary size and transfer effects for Amharic, Hausa, and Swahili @MasakhaneNLP #Masakhane https://t.co/CcB8FWxXSu

Shared by Jimmy Lin at 4/25/2022

Differentiable Search Indices showed that autoregressive transformers can operate as “search engines” to some extent, but not for full scale corpora. FM-Indices change this! Really enjoyed this work led by the amazing @MicheleBevila20!

Shared by Sebastian Riedel at 4/25/2022

Visit our ICLR poster "Measuring the Interpretability of Unsupervised Representations via Quantized Reversed Probing" with @irolaina and A. Vedaldi, from @Oxford_VGG. We linear-probe SSL models, but in ǝsǝʌǝɹ!🤯 For better interpretability. in 1h: https://t.co/99pe94QKur

Shared by Yuki at 4/25/2022

The Role of Permutation Invariance in Linear Mode Connectivity of Neural Networks with @rahiment @osaukh @bneyshabur Poster Session 1 Monday Paper: https://t.co/SDjTlE0nju Poster & video: https://t.co/hSeMBW5rbq Twitter thread: https://t.co/EwDnaQDCWR (2/4)

Shared by Hanie Sedghi at 4/25/2022

Classic active learning problems become more interesting when combined w/ the few-shot abilities of foundation models: in tiny datasets, the kind of data you train on matters a lot. Check out our work (w/ @AlexTamkin, Salil Deshpande, Dat Nguyen, Noah Goodman) towards this end!

Shared by Jesse Mu at 4/24/2022