# The Stanford AI Lab Blog The Stanford AI Lab (SAIL) Blog is a place for SAIL students, faculty, and researchers to share our work with the general public. This is one page of public article previews, not the complete archive. Follow Next page to continue. Summaries are not the original full articles. ## LinkBERT: Improving Language Model Training with Document Link DevFeed: [LinkBERT: Improving Language Model Training with Document Link]() Original publisher: [Read original article]() Author: A Href; Michihiro Yasunaga; Jure Leskovec; Percy Liang Published: 2022-05-31T07:00:00Z Content type: article Language: en Sources: [The Stanford AI Lab Blog]() Topics: [Large Language Model](), [Natural language processing](), [Training AI Models](), [Self-supervised learning](), [Web](), [data](), [Fine-tuning]() Tags: [data](), [google](), [language-models](), [model](), [model-training](), [natural-language-processing](), [self-supervised-learning](), [training](), [web]() ### AI overview The article introduces LinkBERT in the context of language-model pretraining. It explains how BERT and GPT use self-supervised objectives to learn from large text corpora, then describes a limitation of treating documents independently: hyperlinks and citation links connect knowledge across documents and can support multi-hop understanding. The supplied text ends while introducing this challenge. ### Source excerpt Language Model Pretraining Language models (LMs), like BERT 1 and the GPT series 2, achieve remarkable performance on many natural language processing (NLP) tasks. They are now the foundation of today's NLP systems. 3 These models serve important roles in products and tools that we use every day, such as search engines like Google 4 and personal assistants like Alexa 5. These LMs are powerful because they can be pretrained via self-supervised learning on massive amounts of text data on the web without the need for labels, after which the pretrained models can be quickly adapted to a wide range of new tasks without much task-specific finetuning. For instance, BERT is pretrained to predict randomly masked words in original text (masked language modeling), e.g. predicting the masked word "dog" from "My __ is fetching the ball". GPTs are pretrained to predict the next word given a previous sequence of text (causal language modeling), e.g. predicting the next word "ball" from "My dog is fetching the". In either cases, through pretraining, LMs learn to encode various knowledge from a text corpus that helps to perform downstream applications involving language understanding or generation. In particular, LMs can learn world knowledge (associations between concepts like "dog", "fetch", "ball") from training text where the concepts appear together, and help for knowledge-intensive applications like question answering. 6 Challenges. A challenge with most common LM pretraining strategies is that they model a single document at a time. That is, one would split a text corpus into a list of documents and draw training instances for LMs from each document independently. Treating each document independently may pose limitations because documents often have rich dependencies with each other. For instance, text from the web 7 or scientific literature 8 is often used for LM training, but they all have document links, such as hyperlinks and citation links. Document links are crucial bec ## Stanford AI Lab Papers and Talks at ACL 2022 DevFeed: [Stanford AI Lab Papers and Talks at ACL 2022]() Original publisher: [Read original article]() Author: Compiled by Drew A. Hudson Published: 2022-05-25T07:00:00Z Content type: article Language: en Sources: [The Stanford AI Lab Blog]() Topics: [Language models](), [Training AI Models]() Tags: [ai](), [analysis](), [embedding](), [evaluation](), [generation](), [language-models](), [large-language-models](), [retrieval](), [text-generation](), [training]() ### AI overview A Stanford AI Lab roundup of papers, talks, and related material presented at ACL 2022. The listed work covers language-model pretraining, large-language-model analysis, embedding similarity, text generation and summarization evaluation, entity retrieval, dialogue agents, and human-robot interaction. ### Source excerpt The 60th Annual Meeting of the Association for Computational Linguistics (ACL) 2022 is taking place May 22nd - May 27th. We're excited to share all the work from SAIL that's being presented, and you'll find links to papers, videos and blogs below. Feel free to reach out to the contact authors directly to learn more about the work that's happening at Stanford! List of Accepted Papers LinkBERT: Pretraining Language Models with Document Links Authors: Michihiro Yasunaga, Jure Leskovec*, Percy Liang* Contact: myasu@cs.stanford.edu Links: Paper | Website Keywords: language model, pretraining, knowledge, hyperlink, bionlp When classifying grammatical role, BERT doesn't care about word order... except when it matters Authors: Isabel Papadimitriou, Richard Futrell, Kyle Mahowald Contact: isabelvp@stanford.edu Links: Paper Keywords: large language models, analysis, word order, order invariance, grammatical role, syntax, semantics Problems with Cosine as a Measure of Embedding Similarity for High Frequency Words Authors: Kaitlyn Zhou, Kawin Ethayarajh, Dallas Card, Dan Jurafsky Contact: katezhou@stanford.edu Keywords: cosine similarity, training data frequency, model analysis Faithful or Extractive? On Mitigating the Faithfulness-Abstractiveness Trade-off in Abstractive Summarization Authors: Faisal Ladhak, Esin Durmus, He He, Claire Cardie, Kathleen McKeown Contact: esdurmus@stanford.edu Links: Paper Keywords: text summarization, text generation, evaluation, faithfulness Spurious Correlations in Reference-Free Evaluation of Text Generation Authors: Esin Durmus, Faisal Ladhak, Tatsunori Hashimoto Contact: esdurmus@stanford.edu Links: Paper Keywords: text summarization, text generation, dialogue generation, evaluation, metrics, TABi: Type-Aware Bi-Encoders for Open-Domain Entity Retrieval Authors: Megan Leszczynski, Daniel Y. Fu, Mayee F. Chen, Christopher Ré Contact: mleszczy@stanford.edu Links: Paper | Blog Post | Website Keywords: entity retrieval, contrastive learning, bi-en ## Stanford AI Lab Papers and Talks at ICLR 2022 DevFeed: [Stanford AI Lab Papers and Talks at ICLR 2022]() Original publisher: [Read original article]() Author: Compiled by Drew A. Hudson Published: 2022-04-25T07:00:00Z Content type: article Language: en Sources: [The Stanford AI Lab Blog]() Topics: [Artificial Intelligence](), [Benchmark](), [benchmarking](), [Reinforcement learning](), [dataset](), [Language models](), [Graphs](), [Robotics](), [Robotic manipulation]() Tags: [benchmark](), [benchmarking](), [datasets](), [iclr](), [knowledge-graph](), [language-models](), [neural-networks](), [reinforcement-learning](), [robotic-manipulation](), [robotics]() ### AI overview Stanford AI Lab presents its work at ICLR 2022, including papers and talks on reinforcement learning, benchmark datasets, distribution shifts, in-context learning, language models, graph reasoning, model editing, robotics, and robotic manipulation. ### Source excerpt The International Conference on Learning Representations (ICLR) 2022 is being hosted virtually from April 25th - April 29th. We're excited to share all the work from SAIL that's being presented, and you'll find links to papers, videos and blogs below. Feel free to reach out to the contact authors directly to learn more about the work that's happening at Stanford! List of Accepted Papers Autonomous Reinforcement Learning: Formalism and Benchmarking Authors: Archit Sharma*, Kelvin Xu*, Nikhil Sardana, Abhishek Gupta, Karol Hausman, Sergey Levine, Chelsea Finn Contact: architsh@stanford.edu Links: Paper | Website Keywords: reinforcement learning, continual learning, reset-free reinforcement learning MetaShift: A Dataset of Datasets for Evaluating Contextual Distribution Shifts and Training Conflicts Authors: Weixin Liang, James Zou Contact: wxliang@stanford.edu Links: Paper | Video | Website Keywords: benchmark dataset, distribution shift, out-of-domain generalization An Explanation of In-context Learning as Implicit Bayesian Inference Authors: Sang Michael Xie, Aditi Raghunathan, Percy Liang, Tengyu Ma Contact: xie@cs.stanford.edu Links: Paper | Video Keywords: gpt-3, in-context learning, pretraining, few-shot learning GreaseLM: Graph REASoning Enhanced Language Models for Question Answering Authors: Xikun Zhang, Antoine Bosselut, Michihiro Yasunaga, Hongyu Ren, Percy Liang, Christopher D. Manning, Jure Leskovec Contact: xikunz2@cs.stanford.edu Award nominations: Spotlight Links: Paper | Website Keywords: knowledge graph, question answering, language model, commonsense reasoning, graph neural networks, biomedical qa Fast Model Editing at Scale Authors: Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, Christopher D. Manning Contact: eric.mitchell@cs.stanford.edu Links: Paper | Website Keywords: model editing; meta-learning; language models; continual learning; temporal generalization Vision-Based Manipulators Need to Also See from Their Hands Authors: Kyle H ## Discovering the systematic errors made by machine learning models DevFeed: [Discovering the systematic errors made by machine learning models]() Original publisher: [Read original article]() Author: Sabri Eyuboglu, Maya Varma, Khaled Saab, Jared Dunnmon, James Zou and Chris Ré Published: 2022-04-07T07:00:00Z Content type: article Language: en Sources: [The Stanford AI Lab Blog]() Topics: [Machine Learning & Artificial Intelligence](), [Embeddings](), [data](), [model-deployment](), [debug](), [Optimization]() Tags: [data](), [datasets](), [debug](), [embeddings](), [evaluation](), [machine-learning](), [model](), [model-deployment](), [optimization](), [validation]() ### AI overview This article introduces Domino, a method for discovering systematic errors that machine learning models make on coherent, underperforming slices of validation data. It explains how slice awareness supports model evaluation, deployment decisions, debugging, robustness improvements, and safety and fairness analysis, including through cross-modal embeddings. ### Source excerpt Discovering systematic errors with cross-modal embeddings In this blog post, we introduce Domino, a new approach for discovering systematic errors made by machine learning models. We also discuss a framework for quantitatively evaluating methods like Domino. Links: 📄 Paper (ICLR 2022) 🌍 Longer Walkthrough 💻 GitHub 📘 Docs 📒 Google Colab Machine learning models that achieve high overall accuracy often make systematic errors on coherent slices of validation data. What is a slice? A slice is a set of data samples that share a common characteristic. As an example, in large image datasets, photos of vintage cars comprise a slice (i.e. all images in the slice share a common subject). The term slice has a number of synonyms that you might be more familiar with (e.g. subgroup, subpopulation, stratum). These terms are largely interchangeable, but we'll stick with "slice" throughout this post. We say that a model underperforms on a slice if performance on the data samples in the slice is significantly worse than its overall performance. The search for underperforming slices is a critical, but often overlooked, part of model evaluation. When practitioners are aware of the slices on which their models underperform, they can make more informed decisions around model deployment. This is particularly important in safety-critical settings like medicine: a diagnostic model that underperforms on younger patients should likely not be deployed at a pediatric hospital. Slice awareness can also help practitioners debug and improve models: after an underperforming slice is identified, we can improve model robustness by either updating the training dataset or using robust optimization techniques (e.g. Sohoni et al., 2020; Sagawa et al., 2020). Deploying models that underperform on critical data slices may have significant safety or fairness consequences. For example, models trained to detect collapsed lungs in chest X-rays have been shown to make predictions based on the presence of chest d ## Grading Complex Interactive Coding Programs with Reinforcement Learning DevFeed: [Grading Complex Interactive Coding Programs with Reinforcement Learning]() Original publisher: [Read original article]() Author: A Href; Allen Nie; Emma Brunskill; Chris Piech Published: 2022-03-28T07:00:00Z Content type: article Language: en Sources: [The Stanford AI Lab Blog]() Topics: [rlvr](), [Code Challenge](), [Artificial Intelligence](), [Programming](), [NeurIPS](), [browser]() Tags: [agents](), [algorithms](), [browser](), [coding](), [courses](), [games](), [machine-learning](), [neurips](), [programming](), [reinforcement-learning](), [speed]() ### AI overview This article presents the Play to Grade Challenge, which applies reinforcement-learning methods for game-playing AI agents to the automated grading of complex interactive coding assignments. It explains why games and interactive applications are difficult to evaluate automatically and describes the challenge introduced in a NeurIPS 2021 paper. ### Source excerpt [Summary] tl;dr: A tremendous amount of effort has been poured into training AI algorithms to competitively play games that computers have traditionally had trouble with, such as the retro games published by Atari, Go, DotA, and StarCraft II. The practical machine learning knowledge accumulated in developing these algorithms has paved the way for people to now routinely train game-playing AI agents for many games. Following this line of work, we focus on a specific category of games - those developed by students as part of a programming assignment. Can the same algorithms that master Atari games help us grade these game assignments? In our recent NeurIPS 2021 paper, we illustrate the challenges in treating interactive coding assignment grading as game playing and introduce the Play to Grade Challenge. Introduction Massive Online Coding Education has reached striking success over the past decade. Fast internet speed, improved UI design, code editors that are embedded in a browser window allow educational platforms such as Code.org to build a diverse set of courses tailored towards students of different coding experiences and interest levels (for example, Code.org offers "Star War-themed coding challenge," and "Elsa/Frozen themed for-loop writing"). As a non-profit organization, Code.org claims to have reached over 60 million learners across the world 1. Such organizations typically provide a variety of carefully constructed teaching materials such as videos and programming challenges. A challenge faced by these platforms is that of grading assignments. It is well known that grading is critical to student learning 2, in part because it motivates students to complete their assignments. Sometimes manual grading can be feasible in small settings, or automated grading used in simple settings such as when assignments are multiple choice or adopt a fill-in-the-blink modular coding structure. Unfortunately, many of the most exciting assignments, such as developing games or i ## Understanding Deep Learning Algorithms that Leverage Unlabeled Data, Part 1: Self-training DevFeed: [Understanding Deep Learning Algorithms that Leverage Unlabeled Data, Part 1: Self-training]() Original publisher: [Read original article]() Author: A Href; Colin Wei; Jeff Z HaoChen; Tengyu Ma Published: 2022-02-24T08:00:00Z Content type: article Language: en Sources: [The Stanford AI Lab Blog]() Topics: [Deep neural networks](), [Algorithms, Complexity](), [datasets](), [Representation learning](), [ImageNet]() Tags: [algorithms](), [analysis](), [data](), [datasets](), [deep-learning](), [image](), [learning](), [model](), [predictions](), [research](), [training]() ### AI overview This article explains self-training, a semi-supervised learning method that uses a classifier to generate pseudo-labels for unlabeled data and then retrains a new classifier on those labels. It introduces the method and presents a theoretical analysis of why retraining can improve accuracy over the original pseudo-labeler. ### Source excerpt Deep models require a lot of training examples, but labeled data is difficult to obtain. This motivates an important line of research on leveraging unlabeled data, which is often more readily available. For example, large quantities of unlabeled image data can be obtained by crawling the web, whereas labeled datasets such as ImageNet require expensive labeling procedures. In recent empirical developments, models trained with unlabeled data have begun to approach fully-supervised performance (e.g., Chen et al., 2020, Sohn et al., 2020). This series of blog posts will discuss our theoretical work which seeks to analyze recent empirical methods which use unlabeled data. In this first post, we'll analyze self-training, which is a very impactful algorithmic paradigm for semi-supervised learning and domain adaptation. In Part 2, we will use related theoretical ideas to analyze self-supervised contrastive learning algorithms, which have been very effective for unsupervised representation learning. Background: self-training We will first provide a basic overview of self-training algorithms, which are the main focus of this blog post. The core idea is to use some pre-existing classifier \(F_{pl}\) (referred to as the "pseudo-labeler") to make predictions (referred to as "pseudo-labels") on a large unlabeled dataset, and then retrain a new model with the pseudo-labels. For example, in semi-supervised learning, the pseudo-labeler is obtained from training on a small labeled dataset, and is then used to predict pseudo-labels on a larger unlabeled dataset. A new classifier \(F\) is then retrained from scratch to fit the pseudo-labels, using additional regularization. In practice, \(F\) will often be more accurate than the original pseudo-labeler \(F_{pl}\) (Lee 2013). The self-training procedure is depicted below. It is quite surprising that self-training can work so well in practice, given that we retrain on our own predictions, i.e. the pseudo-labels, but not the true labels. ## Stanford AI Lab Papers and Talks at AAAI 2022 DevFeed: [Stanford AI Lab Papers and Talks at AAAI 2022]() Original publisher: [Read original article]() Author: Compiled by Drew A. Hudson Published: 2022-02-22T08:00:00Z Content type: article Language: en Sources: [The Stanford AI Lab Blog]() Topics: [AI, ML & Data Engineering]() Tags: [ai](), [computer-vision](), [conference](), [fact-verification](), [reinforcement-learning]() ### AI overview A Stanford AI Lab overview of accepted AAAI 2022 papers, with links to related papers, videos, websites, and a blog post. The listed work covers cooperative bandits, reinforcement learning, satellite-image object counting, multiagent reinforcement learning, fact verification, and active learning. ### Source excerpt The 36th AAAI Conference on Artificial Intelligence (AAAI 2022) is being hosted virtually from February 22th - March 1st. We're excited to share all the work from SAIL that's being presented, and you'll find links to papers, videos and blogs below. Feel free to reach out to the contact authors directly to learn more about the work that's happening at Stanford. List of Accepted Papers Partner-Aware Algorithms in Decentralized Cooperative Bandit Teams Authors: Erdem Bıyık, Anusha Lalitha, Rajarshi Saha, Andrea Goldsmith, Dorsa Sadigh Contact: ebiyik@stanford.edu Links: Paper | Video | 2nd Video | Website Keywords: bandits, multi-agent systems, collaboration, human-robot interaction, partner-awareness Constraint Sampling Reinforcement Learning: Incorporating Expertise For Faster Learning Authors: Tong Mu, Georgios Theocharous, David Arbour, Emma Brunskill Contact: tongm@stanford.edu Links: Paper Keywords: reinforcement learning, constraints IS-Count: Large-scale Object Counting from Satellite Images with Covariate-based Importance Sampling Authors: Chenlin Meng*, Enci Liu*, Willie Neiswanger, Jiaming Song, Marshall Burke, David Lobell, Stefano Ermon Contact: jesslec@stanford.edu Award nominations: Oral presentation Links: Paper | Blog Post | Website Keywords: remote sensing, sampling PantheonRL Authors: Bidipta Sarkar, Aditi Talati, Andy Shih, Dorsa Sadigh Contact: bidiptas@stanford.edu Links: Paper | Video | Website Keywords: multiagent reinforcement learning; software package; web user interface; adaptive marl; dynamic training interactions Synthetic Disinformation Attacks on Automated Fact Verification Systems Authors: Yibing Du, Antoine Bosselut, Christopher D Manning Contact: antoineb@cs.stanford.edu Links: Paper Keywords: fact checking, fact verification, disinformation, synthetic text Similarity Search for Efficient Active Learning and Search of Rare Concepts Authors: Cody Coleman, Edward Chou, Julian Katz-Samuels, Sean Culatana, Peter Bailis, Alexander C. Berg, ## How to Improve User Experience (and Behavior): Three Papers from Stanford's Alexa Prize Team DevFeed: [How to Improve User Experience (and Behavior): Three Papers from Stanford's Alexa Prize Team]() Original publisher: [Read original article]() Author: A Href; Amelia Hardy; Haojun Li; Abigail See Published: 2022-02-01T08:00:00Z Content type: article Language: en Sources: [The Stanford AI Lab Blog]() Topics: [User Experience](), [Bot](), [Chat Bot](), [Language models](), [Open Source]() Tags: [alexa](), [chat](), [code](), [developers](), [experience](), [language-models](), [modular](), [neural](), [scripted]() ### AI overview This article presents research from Stanford's Alexa Prize team on improving user experience and behavior in socialbot conversations. It discusses user dissatisfaction, responses to offensive behavior, and how conversational control can be balanced between users and bots. The Chirpy Cardinal socialbot uses a modular combination of neural generation and scripted dialogue to support broad, open-domain conversations. ### Source excerpt Introduction In 2019, Stanford entered the Alexa Prize Socialbot Grand Challenge 3 for the first time, with its bot Chirpy Cardinal, which went on to win 2nd place in the competition. In our previous post, we discussed the technical structure of our socialbot and how developers can use our open-source code to develop their own. In this post we share further research conducted while developing Chirpy Cardinal to discover common pain points that users encounter when interacting with socialbots, and strategies for addressing them. The Alexa Prize is a unique research setting, as it allows researchers to study how users interact with a bot when doing so solely for their own motivations. During the competition, US-based Alexa users can say the phrase "let's chat" to speak in English to an anonymous and randomly-selected competing bot. They are free to end the conversation at any time. Since Alexa Prize socialbots are intended to create as natural an experience as possible, they should be capable of long, open-domain social conversations with high coverage of topics. We observed that Chirpy users were interested in many different subjects, from current events (e.g., the coronavirus) to pop culture (e.g., the movie Frozen 2) to personal interests (e.g,. their pets). Chirpy achieves its coverage of these diverse topics by using a modular design that combines both neural generation and scripted dialogue, as described in our previous post. We used this setting to study three questions about socialbot conversations: What do users complain about, and how can we learn from the complaints to improve neurally generated dialogue? What strategies are effective and ineffective in handling and deterring offensive user behavior? How can we shift the balance of power, such that both users and the bot are meaningfully controlling the conversation? We've published papers on each of these topics at SIGDIAL 2021 and in this post, we'll share key findings which provide practical insights for ## Reward Isn't Free: Supervising Robot Learning with Language and Video from the Web DevFeed: [Reward Isn't Free: Supervising Robot Learning with Language and Video from the Web]() Original publisher: [Read original article]() Author: A Href; Suraj Nair Published: 2022-01-21T08:00:00Z Content type: article Language: en Sources: [The Stanford AI Lab Blog]() Topics: [Robotics](), [Deep learning](), [datasets](), [foundation-models](), [Reinforcement learning]() Tags: [datasets](), [deep-learning](), [foundation-models](), [nlp](), [reinforcement-learning](), [robotics](), [vision]() ### AI overview The article examines the challenge of teaching robots to generalize across new environments, tasks, and objects. It discusses using large, diverse datasets and scalable offline learning, while highlighting the limitations of imitation learning and the difficulty of designing or manually annotating reward functions for offline reinforcement learning. ### Source excerpt This work was conducted as part of SAIL and CRFM. Deep learning has enabled improvements in the capabilities of robots on a range of problems such as grasping 1 and locomotion 2 in recent years. However, building the quintessential home robot that can perform a range of interactive tasks, from cooking to cleaning, in novel environments has remained elusive. While a number of hardware and software challenges remain, a necessary component is robots that can generalize their prior knowledge to new environments, tasks, and objects in a zero or few shot manner. For example, a home robot tasked with setting the dining table cannot afford lengthy re-training for every new dish, piece of cutlery, or dining room it may need to interact with. A natural way to enable such generalization in our robots is to train them on rich data sources that contain a wide range of different environments, tasks, and objects. Indeed, this recipe of massive, diverse datasets combined with scalable offline learning algorithms (e.g. self-supervised or cheaply supervised learning) has been the backbone of the many recent successes of foundation models 3 in NLP 456789 and vision 101112. Replicating these impressive generalization and adaptation capabilities in robot learning algorithms would certainly be a step toward robots that can be used in unstructured, real world environments. However, directly extending this recipe to robotics is nontrivial, as we neither have sufficiently large and diverse datasets of robot interaction, nor is it obvious what type of supervision can enable us to scalably learn useful skills from these datasets. On one hand, the popular imitation learning approach relies on expert data which can be expensive to obtain at scale. On the other hand, offline reinforcement learning, which can be performed using non-expert and autonomously-collected data, requires us to define a suitable reward function. Hard-coded reward functions are often task-specific and difficult to design, ## BanditPAM: Almost Linear-Time k-medoids Clustering via Multi-Armed Bandits DevFeed: [BanditPAM: Almost Linear-Time k-medoids Clustering via Multi-Armed Bandits]() Original publisher: [Read original article]() Author: A Href; Mo Tiwari Published: 2021-12-17T08:00:00Z Content type: article Language: en Sources: [The Stanford AI Lab Blog]() Topics: [AI, ML & Data Engineering](), [C++]() Tags: [algorithms](), [c-plus-plus](), [code](), [complexity](), [neurips](), [speed]() ### AI overview BanditPAM is a publicly available k-medoids clustering implementation that reduces the stated runtime from O(n²) to O(n log n). It is pip-installable, written in C++, and designed to resemble the sklearn.cluster.KMeans interface. ### Source excerpt TL;DR Want something better than \(k\)-means? Our state-of-the-art \(k\)-medoids algorithm from NeurIPS, BanditPAM, is now publicly available! \(\texttt{pip install banditpam}\) and you're good to go! Like the \(k\)-means problem, the \(k\)-medoids problem is a clustering problem in which our objective is to partition a dataset into disjoint subsets. In \(k\)-medoids, however, we require that the cluster centers must be actual datapoints, which permits greater interpretability of the cluster centers. \(k\)-medoids also works better with arbitrary distance metrics, so your clustering can be more robust to outliers if you're using metrics like \(L_1\). Despite these advantages, most people don't use \(k\)-medoids because prior algorithms were too slow. In our NeurIPS paper, BanditPAM, we sped up the best known algorithm from \(O(n^2)\) to \(O(n\text{log}n)\). We've released our implementation, which is pip-installable. It's written in C++ for speed and supports parallelization and intelligent caching, at no extra complexity to end users. Its interface also matches the \(\texttt{sklearn.cluster.KMeans}\) interface, so minimal changes are necessary to existing code. Useful Links: 3-minute video summary PyPI Github Repository Full Paper \(k\)-means vs. \(k\)-medoids If you're an ML practitioner, you're probably familiar with the \(k\)-means problem. In fact, you may know some of the common algorithms for the \(k\)-means problem. You're much less likely, however, familiar with the \(k\)-medoids problem. The \(k\)-medoids problem is a clustering problem similar to \(k\)-means. Given a dataset, we want to partition our dataset into subsets where the points in each cluster are closer to a single cluster center than all other \(k-1\) cluster centers. Unlike in \(k\)-means, however, the \(k\)-medoids problem requires cluster centers to be actual datapoints. Figure 1: The \(k\)-medoids solution (left) forces the cluster centers to be actual datapoints. This solution is often di ## Stanford AI Lab Papers and Talks at NeurIPS 2021 DevFeed: [Stanford AI Lab Papers and Talks at NeurIPS 2021]() Original publisher: [Read original article]() Author: Compiled by Drew A. Hudson Published: 2021-12-06T08:00:00Z Content type: article Language: en Sources: [The Stanford AI Lab Blog]() Topics: [NeurIPS](), [Artificial Intelligence](), [Reverse Engineering](), [data](), [Reinforcement learning]() Tags: [ai](), [generative](), [imitation](), [learning](), [models](), [neural-networks](), [neurips](), [reinforcement-learning](), [reverse-engineering](), [transformers]() ### AI overview Stanford AI Lab presents its research at NeurIPS 2021, including work on generative models, recurrent and transformer-based architectures, state-space models, emergent communication, reinforcement learning, imitation learning, and neural coding. ### Source excerpt The thirty-fifth Conference on Neural Information Processing Systems (NeurIPS) 2021 is being hosted virtually from Dec 6th - 14th. We're excited to share all the work from SAIL that's being presented at the main conference, at the Datasets and Benchmarks track and the various workshops, and you'll find links to papers, videos and blogs below. Some of the members in our SAIL community also serve as co-organizers of several exciting workshops that will take place on Dec 13-14, so we hope you will check them out! Feel free to reach out to the contact authors and the workshop organizers directly to learn more about the work that's happening at Stanford! Main Conference Improving Compositionality of Neural Networks by Decoding Representations to Inputs Authors: Mike Wu, Noah Goodman, Stefano Ermon Contact: wumike@stanford.edu Links: Paper Keywords: generative models, compositionality, decoder Reverse engineering recurrent neural networks with Jacobian switching linear dynamical systems Authors: Jimmy T.H. Smith, Scott W. Linderman, David Sussillo Contact: jsmith14@stanford.edu Links: Paper | Website Keywords: recurrent neural networks, switching linear dynamical systems, interpretability, fixed points Compositional Transformers for Scene Generation Authors: Drew A. Hudson, C. Lawrence Zitnick Contact: dorarad@cs.stanford.edu Links: Paper | Github Keywords: GANs, transformers, compositionality, scene synthesis Combining Recurrent, Convolutional, and Continuous-time Models with Linear State Space Layers Authors: Albert Gu, Isys Johnson, Karan Goel, Khaled Saab, Tri Dao, Atri Rudra, Chris Ré Contact: albertgu@stanford.edu Links: Paper Keywords: recurrent neural networks, rnn, continuous models, state space, long range dependencies, sequence modeling Emergent Communication of Generalizations Authors: Jesse Mu, Noah Goodman Contact: muj@stanford.edu Links: Paper | Video Keywords: emergent communication, multi-agent communication, language grounding, compositionality Deep Lear ## Stanford AI Lab Papers at CoRL 2021 DevFeed: [Stanford AI Lab Papers at CoRL 2021]() Original publisher: [Read original article]() Author: Compiled by Drew A. Hudson Published: 2021-11-05T07:00:00Z Content type: article Language: en Sources: [The Stanford AI Lab Blog]() Topics: [Reinforcement learning](), [Artificial Intelligence](), [benchmarking](), [Robotic manipulation](), [Benchmark]() Tags: [ai](), [benchmark](), [benchmarking](), [imitation](), [learning](), [offline](), [reinforcement-learning](), [rl](), [robotic-manipulation]() ### AI overview The Stanford AI Lab presents its work accepted at the Conference on Robot Learning (CoRL 2021). The article lists projects covering language-informed robot autonomy, household-activity benchmarking, human-robot collaboration, imitation learning, model-based and deep reinforcement learning, reward and preference learning, and robotic manipulation. ### Source excerpt The Conference on Robot Learning (CoRL 2021) will take place next week. We're excited to share all the work from SAIL that will be presented, and you'll find links to papers, videos and blogs below. Feel free to reach out to the contact authors directly to learn more about the work that's happening at Stanford! List of Accepted Papers LILA: Language-Informed Latent Actions Authors: Siddharth Karamcheti*, Megha Srivastava*, Percy Liang, Dorsa Sadigh Contact: skaramcheti@cs.stanford.edu, megha@cs.stanford.edu Keywords: natural language, shared autonomy, human-robot interaction BEHAVIOR: Benchmark for Everyday Household Activities in Virtual, Interactive, and Ecological Environments Authors: Sanjana Srivastava*, Chengshu Li*, Michael Lingelbach*, Roberto Martín-Martín*, Fei Xia, Kent Vainio, Zheng Lian, Cem Gokmen, Shyamal Buch, C. Karen Liu, Silvio Savarese, Hyowon Gweon, Jiajun Wu, Li Fei-Fei Contact: sanjana2@stanford.edu Links: Paper | Website Keywords: embodied ai, benchmarking, household activities Co-GAIL: Learning Diverse Strategies for Human-Robot Collaboration Authors: Chen Wang, Claudia Pérez-D'Arpino, Danfei Xu, Li Fei-Fei, C. Karen Liu, Silvio Savarese Contact: chenwj@stanford.edu Links: Paper | Website Keywords: learning for human-robot collaboration, imitation learning DiffImpact: Differentiable Rendering and Identification of Impact Sounds Authors: Samuel Clarke, Negin Heravi, Mark Rau, Ruohan Gao, Jiajun Wu, Doug James, Jeannette Bohg Contact: spclarke@stanford.edu Links: Paper | Website Keywords: differentiable sound rendering, auditory scene analysis Example-Driven Model-Based Reinforcement Learning for Solving Long-Horizon Visuomotor Tasks Authors: Bohan Wu, Suraj Nair, Li Fei-Fei*, Chelsea Finn* Contact: bohanwu@cs.stanford.edu Links: Paper Keywords: model-based reinforcement learning, long-horizon tasks GRAC: Self-Guided and Self-Regularized Actor-Critic Authors: Lin Shao, Yifan You, Mengyuan Yan, Shenli Yuan, Qingyun Sun, Jeannette Bohg Contact: ## Stanford AI Lab Papers at EMNLP/CoNLL 2021 DevFeed: [Stanford AI Lab Papers at EMNLP/CoNLL 2021]() Original publisher: [Read original article]() Author: Compiled by Drew A. Hudson Published: 2021-11-05T07:00:00Z Content type: article Language: en Sources: [The Stanford AI Lab Blog]() Topics: [Natural language processing](), [Large Language Model](), [dataset](), [Reinforcement learning](), [Algorithms, Complexity]() Tags: [ai](), [blog-post](), [complexity](), [data](), [language-models](), [nlp](), [reinforcement-learning]() ### AI overview Stanford AI Lab presents its accepted papers and related work for EMNLP 2021 and CoNLL 2021. The collection covers natural language processing topics including language generation, communication-based training, biomedical named entity disambiguation, contract-focused natural language inference, emergent communication, language models, grammatical error correction, multilingual machine translation, and legal NLP. ### Source excerpt The 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP 2021) will take place next week, colocated with CoNLL 2021. We're excited to share all the work from SAIL that will be presented, and you'll find links to papers, videos and blogs below. Feel free to reach out to the contact authors directly to learn more about the work that's happening at Stanford! List of Accepted Papers Calibrate your listeners! Robust communication-based training for pragmatic speakers Authors: Rose E. Wang, Julia White, Jesse Mu, Noah D. Goodman Contact: rewang@stanford.edu Links: Paper | Video Keywords: language generation, pragmatics, communication-based training, calibration, uncertainty Cross-Domain Data Integration for Named Entity Disambiguation in Biomedical Text Authors: Maya Varma, Laurel Orr, Sen Wu, Megan Leszczynski, Xiao Ling, Christopher Ré Contact: mvarma2@stanford.edu Links: Paper | Video Keywords: named entity disambiguation, biomedical text, rare entities, data integration ContractNLI: A Dataset for Document-level Natural Language Inference for Contracts Authors: Yuta Koreeda, Christopher D. Manning Contact: koreeda@stanford.edu Links: Paper | Website Keywords: natural language inference, contract, law, legal, dataset Venue: The Findings of EMNLP 2021 The Emergence of the Shape Bias Results from Communicative Efficiency Authors: Eva Portelance, Michael C. Frank, Dan Jurafsky, Alessandro Sordoni, Romain Laroche Contact: portelan@stanford.edu Links: Paper | Website Keywords: emergent communication, shape bias, multi-agent reinforcement learning, language learning, language acquisition Conference: CoNLL LM-Critic: Language Models for Unsupervised Grammatical Error Correction Authors: Michihiro Yasunaga, Jure Leskovec, Percy Liang. Contact: myasu@cs.stanford.edu Links: Paper | Blog Post | Website Keywords: language model, grammatical error correction, unsupervised translation Sensitivity as a complexity measure for sequence classification tasks Authors: ## Selective Classification Can Magnify Disparities Across Groups DevFeed: [Selective Classification Can Magnify Disparities Across Groups]() Original publisher: [Read original article]() Author: A Href; Erik Jones Published: 2021-10-13T07:00:00Z Content type: article Language: en Sources: [The Stanford AI Lab Blog]() Topics: [Machine learning](), [model-deployment](), [data](), [datasets](), [Natural language processing]() Tags: [iclr](), [ml](), [model](), [models](), [nlp](), [overview](), [vision]() ### AI overview The article examines selective classification, in which machine-learning models can abstain when uncertain. It shows that although abstention can improve average accuracy, it may fail to help or may worsen accuracy for clinically important subgroups, such as patients with pleural effusion who have not yet received a chest drain. The article discusses the underlying failure mode, theoretical implications, and approaches for building more equitable selective classifiers. ### Source excerpt Selective classification, where models are allowed to "abstain" when they are uncertain about a prediction, is a useful approach for deploying models in settings where errors are costly. For example, in medicine, model errors can have life-or-death ramifications, but abstentions can be easily handled by backing off to a doctor, who then makes a diagnosis. Across a range of applications from vision 123 and NLP 45, even simple selective classifiers, relying only on model logits, routinely and often dramatically improve accuracy by abstaining. This makes selective classification a compelling tool for ML practitioners 67. However, in our recent ICLR paper, we find that despite reliably improving average accuracy, selective classification can fail to improve and even hurt the accuracy over certain subpopulations of the data. As a motivating example, consider the task of diagnosing pleural effusion, or fluid in the lungs, from chest X-rays. Pleural effusion is often treated with a chest drain, so many pleural effusion cases also have chest drains, while most cases without pleural effusion do not have chest drains 8. While selective classification improves average accuracy for this task, we find that it does not appreciably improve accuracy on the most clinically relevant subgroup, or subpopulation, of the data: those that have pleural effusion but don't yet have a chest drain, i.e. those that have pleural effusion but have not yet been treated for it. Practitioners, thus, should be wary of these potential failure modes of using selective classification in the wild. Example of the spurious correlation setup. This patient has a pleural effusion (excess fluid in the lung), but does not yet have a chest drain. The model, relying on the presence of a chest drain to make a prediction, incorrectly predicts negative. To further outline this critical failure mode of selective classification, we'll first provide an overview of selective classification. We then demonstrate empirical ## Stanford AI Lab Papers at ICCV 2021 DevFeed: [Stanford AI Lab Papers at ICCV 2021]() Original publisher: [Read original article]() Author: Compiled by Drew A. Hudson Published: 2021-10-08T07:00:00Z Content type: article Language: en Sources: [The Stanford AI Lab Blog]() Topics: [Artificial Intelligence](), [Self-supervised learning](), [multimodal](), [Representation learning](), [object-detection](), [Transformer]() Tags: [3d](), [ai](), [computer-vision](), [diffusion](), [fairness-accountability-transparency](), [learning](), [multimodal](), [object-detection](), [reasoning](), [video](), [vision]() ### AI overview Stanford AI Lab presents its accepted work for ICCV 2021, including research on multimodal medical image recognition, self-supervised learning, diffusion-based shape generation, object pose tracking, human-object relationships in videos, 3D human motion, computational photography, and related vision topics. ### Source excerpt The International Conference on Computer Vision (ICCV 2021) will be hosted virtually next week. We're excited to share all the work from SAIL that will be presented, and you'll find links to papers, videos and blogs below. Feel free to reach out to the contact authors directly to learn more about the work that's happening at Stanford! List of Accepted Papers GLoRIA: A Multimodal Global-Local Representation Learning Framework for Label-efficient Medical Image Recognition Authors: Mars Huang Contact: mschuang@stanford.edu Keywords: medical image, self-supervised learning, multimodal fusion 3D Shape Generation and Completion Through Point-Voxel Diffusion Authors: Linqi Zhou, Yilun Du, Jiajun Wu Contact: linqizhou@stanford.edu Links: Paper | Video | Website Keywords: diffusion, shape generation CAPTRA: CAtegory-level Pose Tracking for Rigid and Articulated Objects from Point Clouds Authors: Yijia Weng*, He Wang*, Qiang Zhou, Yuzhe Qin, Yueqi Duan, Qingnan Fan, Baoquan Chen, Hao Su, Leonidas J. Guibas Contact: yijiaw@stanford.edu Award nominations: Oral Presentation Links: Paper | Video | Website Keywords: category-level object pose tracking, articulated objects Detecting Human-Object Relationships in Videos Authors: Jingwei Ji, Rishi Desai, Juan Carlos Niebles Contact: jingweij@cs.stanford.edu Links: Paper Keywords: human-object relationships, video, detection, transformer, spatio-temporal reasoning Geography-Aware Self-Supervised Learning Authors: Kumar Ayush, Burak Uzkent, Chenlin Meng, Kumar Tanmay, Marshall Burke, David Lobell, Stefano Ermon Contact: kayush@cs.stanford.edu, chenlin@stanford.edu Links: Paper | Website Keywords: self-supervised learning, contrastive learning, remote sensing, spatio-temporal, classification, object detection, segmentation HuMoR: 3D Human Motion Model for Robust Pose Estimation Authors: Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, Leonidas Guibas Contact: drempe@stanford.edu Award nominations: Oral Presentati