The Stanford AI Lab (SAIL) is making significant contributions to the field of artificial intelligence with a robust lineup of papers and talks presented at the 35th Conference on Neural Information Processing Systems (NeurIPS) 2021, held virtually from December 6th to 14th. The conference features a diverse array of research topics, including advancements in neural networks, reinforcement learning, data efficiency, and causal inference, among others. Below is a concise summary of the key presentations and workshops organized by SAIL members: ### Main Conference Contributions 1. **Improving Compositionality of Neural Networks by Decoding Representations to Inputs** - **Authors:** Mike Wu, Noah Goodman, Stefano Ermon - **Summary:** This paper explores enhancing the compositionality of neural networks by decoding representations back to inputs, a technique that could improve the interpretability and flexibility of these models. 2. **Reverse Engineering Recurrent Neural Networks with Jacobian Switching Linear Dynamical Systems** - **Authors:** Jimmy T.H. Smith, Scott W. Linderman, David Sussillo - **Summary:** The authors introduce a method to reverse engineer recurrent neural networks using Jacobian switching linear dynamical systems, which helps in understanding the internal dynamics and fixed points of these networks, contributing to their interpretability. 3. **Compositional Transformers for Scene Generation** - **Authors:** Drew A. Hudson, C. Lawrence Zitnick - **Summary:** This work proposes the use of Transformers in a compositional manner to generate scenes, leveraging the strengths of GANs and compositionality to create more realistic and diverse outputs. 4. **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 - **Summary:** The paper presents a novel approach to combining different types of neural networks using linear state space layers, which can better handle long-range dependencies and sequence modeling. 5. **Emergent Communication of Generalizations** - **Authors:** Jesse Mu, Noah Goodman - **Summary:** This research investigates how multi-agent systems can develop communication strategies that generalize well, using a combination of language grounding and compositionality. 6. **Deep Learning on a Data Diet: Finding Important Examples Early in Training** - **Authors:** Mansheej Paul, Surya Ganguli, Gintare Karolina Dziugaite - **Summary:** The authors propose a method to identify and prune less important data examples early in the training process, which can lead to more efficient and effective deep learning models. 7. **ELLA: Exploration through Learned Language Abstraction** - **Authors:** Suvir Mirchandani, Siddharth Karamcheti, Dorsa Sadigh - **Summary:** ELLA is a framework that uses learned language abstraction to guide exploration in reinforcement learning tasks, potentially improving the performance and efficiency of these algorithms. 8. **CSDI: Conditional Score-based Diffusion Models for Probabilistic Time Series Imputation** - **Authors:** Yusuke Tashiro, Jiaming Song, Yang Song, Stefano Ermon - **Summary:** CSDI is a new model that uses score-based diffusion techniques to impute missing data in time series, offering a probabilistic approach that can handle complex dependencies. 9. **Confidence-Aware Imitation Learning from Demonstrations with Varying Optimality** - **Authors:** Songyuan Zhang, Zhangjie Cao, Dorsa Sadigh, Yanan Sui - **Summary:** This paper addresses the challenge of learning from suboptimal demonstrations by developing a confidence-aware imitation learning method, which can improve the robustness and performance of AI systems in real-world scenarios. 10. **Explaining heterogeneity in medial entorhinal cortex with task-driven neural networks** - **Authors:** Aran Nayebi, Alexander Attinger, Malcolm G. Campbell, Kiah Hardcastle, Isabel I.C. Low, Caitlin S. Mallory, Gabriel C. Mel, Ben Sorscher, Alex H. Williams, Surya Ganguli, Lisa M. Giocomo, Daniel L.K. Yamins - **Summary:** The research uses task-driven neural networks to explain the heterogeneity observed in the medial entorhinal cortex, a brain region crucial for navigation and spatial memory. 11. **On the theory of reinforcement learning with once-per-episode feedback** - **Authors:** Niladri Chatterji, Aldo Pacchiano, Peter Bartlett, Michael Jordan - **Summary:** This theoretical work explores reinforcement learning with limited feedback, providing insights into how agents can learn effectively with minimal information. 12. **HyperSPNs: Compact and Expressive Probabilistic Circuits** - **Authors:** Andy Shih, Dorsa Sadigh, Stefano Ermon - **Summary:** The paper introduces HyperSPNs, a new type of probabilistic circuit that is both compact and expressive, suitable for various generative modeling tasks. 13. **COMBO: Conservative Offline Model-Based Policy Optimization** - **Authors:** Tianhe Yu, Aviral Kumar, Rafael Rafailov, Aravind Rajeswaran, Sergey Levine, Chelsea Finn - **Summary:** COMBO is a conservative approach to offline model-based policy optimization, designed to improve the reliability and performance of reinforcement learning algorithms when trained on static datasets. 14. **Lossy Compression for Lossless Prediction** - **Authors:** Yann Dubois, Benjamin Bloem-Reddy, Karen Ullrich, Chris J. Maddison - **Summary:** The authors present a method for lossy compression that preserves the ability to make accurate predictions, a technique that could significantly reduce storage and computational costs in machine learning. 15. **Capturing implicit hierarchical structure in 3D biomedical images with self-supervised hyperbolic representations** - **Authors:** Joy Hsu, Jeffrey Gu, Gong-Her Wu, Wah Chiu, Serena Yeung - **Summary:** This research uses self-supervised hyperbolic representations to capture hierarchical structures in 3D biomedical images, enhancing the capabilities of machine learning in medical imaging. 16. **Maximum Likelihood Training of Score-Based Diffusion Models** - **Authors:** Yang Song, Conor Durkan, Iain Murray, Stefano Ermon - **Summary:** The paper discusses the maximum likelihood training of score-based diffusion models, a technique that can improve the generation of high-quality data in generative models. 17. **Safe Reinforcement Learning by Imagining the Near Future** - **Authors:** Garrett Thomas, Yuping Luo, Tengyu Ma - **Summary:** This work introduces a method for safe reinforcement learning by enabling agents to imagine and evaluate potential future outcomes, reducing the risk of unsafe actions. 18. **Pseudo-Spherical Contrastive Divergence** - **Authors:** Lantao Yu, Jiaming Song, Yang Song, Stefano Ermon - **Summary:** The authors propose a new contrastive divergence method for deep generative models, which can improve the training efficiency and model accuracy. 19. **IQ-Learn: Inverse soft-Q Learning for Imitation** - **Authors:** Divyansh Garg, Shuvam Chakraborty, Chris Cundy, Jiaming Song, Stefano Ermon - **Summary:** IQ-Learn is a method that combines inverse reinforcement learning and soft-Q learning to enable efficient imitation learning, particularly useful in complex environments. 20. **Intrinsic Dimension, Persistent Homology and Generalization in Neural Networks** - **Authors:** Tolga Birdal, Aaron Lou, Leonidas Guibas, Umut Simsekli - **Summary:** This paper explores the relationship between intrinsic dimension, persistent homology, and generalization in neural networks, providing theoretical insights into the performance of these models. ### Datasets and Benchmarks Track 1. **ReaSCAN: Compositional Reasoning in Language Grounding** - **Authors:** Zhengxuan Wu, Elisa Kreiss, Desmond Ong, Christopher Potts - **Summary:** ReaSCAN is a new dataset designed to evaluate compositional reasoning in language grounding tasks, pushing the boundaries of NLP and multi-modal learning. 2. **ATOM3D: Tasks on Molecules in Three Dimensions** - **Authors:** Raphael J.L. Townshend, Martin Vögele, Patricia Suriana, Alexander Derry, Alexander S. Powers, Yianni Laloudakis, Sidhika Balachandar, Bowen Jing, Brandon Anderson, Stephan Eismann, Risi Kondor, Russ B. Altman, Ron O. Dror - **Summary:** ATOM3D is a benchmark for evaluating machine learning models on tasks involving 3D molecular structures, which is crucial for drug discovery and materials science. 3. **Dynamic Environments with Deformable Objects** - **Authors:** Rika Antonova, Peiyang Shi, Hang Yin, Zehang Weng, Danica Kragic - **Summary:** This dataset focuses on dynamic environments with deformable objects, challenging machine learning models to handle real-world complexity and variability. 4. **Personalized Benchmarking with the Ludwig Benchmarking Toolkit** - **Authors:** Avanika Narayan, Piero Molino, Karan Goel, Willie Neiswanger, Christopher Ré - **Summary:** The Ludwig Benchmarking Toolkit provides a framework for personalized benchmarking, allowing researchers to evaluate and compare machine learning models more effectively. 5. **SustainBench: Benchmarks for Monitoring the Sustainable Development Goals with Machine Learning** - **Authors:** Christopher Yeh, Chenlin Meng, Sherrie Wang, Anne Driscoll, Erik Rozi, Patrick Liu, Jihyeon Lee, Marshall Burke, David Lobell, Stefano Ermon - **Summary:** SustainBench is a suite of benchmarks designed to monitor the progress of sustainable development goals using machine learning, emphasizing the practical impact of AI in societal challenges. ### Workshops 1. **Machine Learning for Structural Biology Workshop** - **Organizers:** Namrata Anand, Bonnie Berger, Wouter Boomsma, Erika DeBenedictis, Stephan Eismann, John Ingraham, Sergey Ovchinnikov, Roshan Rao, Raphael Townshend, Ellen Zhong - **Summary:** This workshop focuses on the application of machine learning to structural biology, exploring how AI can aid in understanding and predicting the structure and function of biological molecules. 2. **Controllable Generative Modeling in Language and Vision (CtrlGen Workshop)** - **Organizers:** Steven Y. Feng, Drew A. Hudson, Anusha Balakrishnan, Varun Gangal, Dongyeop Kang, Tatsunori Hashimoto, Joel Tetreault - **Summary:** The CtrlGen Workshop discusses advancements in controllable generative modeling, aiming to develop techniques that allow for more precise control over the outputs of generative models in both language and vision tasks. 3. **DistShift Workshop** - **Organizers:** Shiori Sagawa, Pang Wei Koh, Fanny Yang, Hongseok Namkoong, Jiashi Feng, Kate Saenko, Percy Liang, Sarah Bird, Sergey Levine - **Summary:** DistShift addresses the challenges of distribution shifts in machine learning, with a focus on developing methods that can adapt to changing data conditions and improve model robustness. 4. **Data-centric AI Workshop** - **Organizers:** Andrew Ng, Lora Aroyo, Cody Coleman, Greg Diamos, Vijay Janapa Reddi, Joaquin Vanschoren, Carole-Jean Wu, Sharon Zhou - **Summary:** This workshop emphasizes the importance of data in AI, exploring how data-centric approaches can enhance the performance and reliability of machine learning models. 5. **Physical Reasoning and Inductive Biases for the Real World Workshop** - **Organizers:** Krishna Murthy Jatavallabhula, Rika Antonova, Kevin Smith, Hsiao-Yu (Fish) Tung, Florian Shkurti, Jeannette Bohg, Josh Tenenbaum - **Summary:** The workshop explores how physical reasoning and inductive biases can be integrated into AI models to better handle real-world tasks, particularly in robotics and other physical systems. ### Workshop Papers 1. **How Does Contrastive Pre-training Connect Disparate Domains?** - **Authors:** Kendrick Shen, Robbie Jones, Ananya Kumar, Sang Michael Xie, Percy Liang - **Summary:** This paper investigates the mechanisms behind contrastive pre-training and how it can bridge different domains, providing insights into the transferability of learned representations. 2. **Optimal Representations for Covariate Shifts** - **Authors:** Yann Dubois, Yangjun Ruan, Chris J. Maddison - **Summary:** The authors explore the theoretical foundations for optimal representations that can handle covariate shifts, enhancing the robustness of machine learning models in changing environments. 3. **Correct-N-Contrast: a Contrastive Approach for Improving Robustness to Spurious Correlations** - **Authors:** Michael Zhang, Nimit S. Sohoni, Hongyang R. Zhang, Chelsea Finn, Christopher Ré - **Summary:** This paper introduces a contrastive learning method to improve the robustness of models to spurious correlations, a common issue in real-world datasets. 4. **Extending the WILDS Benchmark for Unsupervised Adaptation** - **Authors:** Shiori Sagawa, Pang Wei Koh, Tony Lee, Irena Gao, Sang Michael Xie, Kendrick Shen, Ananya Kumar, Weihua Hu, Michihiro Yasunaga, Henrik Marklund, Sara Beery, Etienne David, Ian Stavness, Wei Guo, Jure Leskovec, Kate Saenko, Tatsunori Hashimoto, Sergey Levine, Chelsea Finn, Percy Liang - **Summary:** The WILDS benchmark is extended to include unsupervised adaptation tasks, providing a more comprehensive evaluation framework for machine learning models in diverse and challenging scenarios. 5. **Sharp Bounds for Federated Averaging (Local SGD) and Continuous Perspective** - **Authors:** Margalit Glasgow, Honglin Yuan, Tengyu Ma - **Summary:** This paper provides sharp theoretical bounds for federated averaging (Local SGD), a key technique in distributed machine learning, and discusses its continuous perspective. 6. **What Matters in Learning from Offline Human Demonstrations for Robot Manipulation** - **Authors:** Ajay Mandlekar, Danfei Xu, Josiah Wong, Soroush Nasiriany, Chen Wang, Rohun Kulkarni, Li Fei-Fei, Silvio Savarese, Yuke Zhu, Roberto Martín-Martín - **Summary:** The authors analyze the factors that influence the effectiveness of learning from offline human demonstrations in robot manipulation tasks, highlighting the importance of diverse and high-quality data. 7. **An Algorithmic Theory of Metacognition in Minds and Machines** - **Authors:** Rylan Schaeffer - **Summary:** This paper develops an algorithmic theory of metacognition, exploring how both biological minds and AI systems can make decisions about their own learning processes. 8. **Beyond Ads: Sequential Decision-Making Algorithms in Public Policy** - **Authors:** Peter Henderson, Ben Chugg, Brandon Anderson, Daniel E. Ho - **Summary:** The research applies sequential decision-making algorithms to public policy, demonstrating their potential in areas beyond traditional advertising, such as healthcare and education. 9. **Tracking Urbanization in Developing Regions with Remote Sensing Spatial-Temporal Super-Resolution** - **Authors:** Yutong He, William Zhang, Chenlin Meng, Marshall Burke, David B. Lobell, Stefano Ermon - **Summary:** This paper uses remote sensing and spatial-temporal super-resolution techniques to track urbanization in developing regions, providing valuable insights for urban planning and policy. 10. **Likelihood-free Density Ratio Acquisition Functions are not Equivalent to Expected Improvements** - **Authors:** Jiaming Song, Stefano Ermon - **Summary:** The authors clarify the theoretical differences between likelihood-free density ratio acquisition functions and expected improvements, contributing to the understanding of Bayesian optimization. 11. **Exploiting Proximity Search and Easy Examples to Select Rare Events** - **Authors:** Daniel Kang, Alex Derhacobian, Kaoru Tsuji, Trevor Hebert, Peter Bailis, Tadashi Fukami, Tatsunori Hashimoto, Yi Sun, Matei Zaharia - **Summary:** This paper proposes a method to select rare events by leveraging proximity search and easy examples, which can improve the efficiency and accuracy of data-driven decision-making processes. ### Conclusion The Stanford AI Lab's participation in NeurIPS 2021 showcases a wide range of cutting-edge research in AI, from theoretical foundations to practical applications. The lab's contributions to neural networks, reinforcement learning, data efficiency, and causal inference highlight the ongoing advancements and innovations in the field. Additionally, the involvement in workshops and the development of new datasets and benchmarks underscore the lab's commitment to fostering collaborative and impactful research. These presentations and discussions are expected to drive further progress and inspire new directions in AI research.