DeepMind的游戏理论与多代理团队研究多学科学习的几个方面，从计算近似值到游戏理论中的基本概念，再到在富裕的空间环境中模拟社会困境，并在困难的团队协调任务中培训3-D类人动物。我们小组的一个签名目的是使用DeepMind在DeepMind中提供的资源和专业知识，以深入强化学习来探索复杂环境中的多代理系统，并使用这些基准来提高我们的理解。在这里，我们总结了我们团队的最新工作，并提出了一种分类法，我们认为这重点介绍了多代理研究中许多重要的开放挑战。

在过去的十年中，多智能经纪人强化学习（Marl）已经有了重大进展，但仍存在许多挑战，例如高样本复杂性和慢趋同稳定的政策，在广泛的部署之前需要克服，这是可能的。然而，在实践中，许多现实世界的环境已经部署了用于生成策略的次优或启发式方法。一个有趣的问题是如何最好地使用这些方法作为顾问，以帮助改善多代理领域的加强学习。在本文中，我们提供了一个原则的框架，用于将动作建议纳入多代理设置中的在线次优顾问。我们描述了在非传记通用随机游戏环境中提供多种智能强化代理（海军上将）的问题，并提出了两种新的基于Q学习的算法：海军上将决策（海军DM）和海军上将 - 顾问评估（Admiral-AE） ，这使我们能够通过适当地纳入顾问（Admiral-DM）的建议来改善学习，并评估顾问（Admiral-AE）的有效性。我们从理论上分析了算法，并在一般加上随机游戏中提供了关于他们学习的定点保证。此外，广泛的实验说明了这些算法：可以在各种环境中使用，具有对其他相关基线的有利相比的性能，可以扩展到大状态行动空间，并且对来自顾问的不良建议具有稳健性。

As language models (LMs) scale, they develop many novel behaviors, good and bad, exacerbating the need to evaluate how they behave. Prior work creates evaluations with crowdwork (which is time-consuming and expensive) or existing data sources (which are not always available). Here, we automatically generate evaluations with LMs. We explore approaches with varying amounts of human effort, from instructing LMs to write yes/no questions to making complex Winogender schemas with multiple stages of LM-based generation and filtering. Crowdworkers rate the examples as highly relevant and agree with 90-100% of labels, sometimes more so than corresponding human-written datasets. We generate 154 datasets and discover new cases of inverse scaling where LMs get worse with size. Larger LMs repeat back a dialog user's preferred answer ("sycophancy") and express greater desire to pursue concerning goals like resource acquisition and goal preservation. We also find some of the first examples of inverse scaling in RL from Human Feedback (RLHF), where more RLHF makes LMs worse. For example, RLHF makes LMs express stronger political views (on gun rights and immigration) and a greater desire to avoid shut down. Overall, LM-written evaluations are high-quality and let us quickly discover many novel LM behaviors.

As AI systems become more capable, we would like to enlist their help to supervise other AIs. We experiment with methods for training a harmless AI assistant through self-improvement, without any human labels identifying harmful outputs. The only human oversight is provided through a list of rules or principles, and so we refer to the method as 'Constitutional AI'. The process involves both a supervised learning and a reinforcement learning phase. In the supervised phase we sample from an initial model, then generate self-critiques and revisions, and then finetune the original model on revised responses. In the RL phase, we sample from the finetuned model, use a model to evaluate which of the two samples is better, and then train a preference model from this dataset of AI preferences. We then train with RL using the preference model as the reward signal, i.e. we use 'RL from AI Feedback' (RLAIF). As a result we are able to train a harmless but non-evasive AI assistant that engages with harmful queries by explaining its objections to them. Both the SL and RL methods can leverage chain-of-thought style reasoning to improve the human-judged performance and transparency of AI decision making. These methods make it possible to control AI behavior more precisely and with far fewer human labels.

As artificial intelligence (AI) becomes a prominent part of modern life, AI literacy is becoming important for all citizens, not just those in technology careers. Previous research in AI education materials has largely focused on the introduction of terminology as well as AI use cases and ethics, but few allow students to learn by creating their own machine learning models. Therefore, there is a need for enriching AI educational tools with more adaptable and flexible platforms for interested educators with any level of technical experience to utilize within their teaching material. As such, we propose the development of an open-source tool (Build-a-Bot) for students and teachers to not only create their own transformer-based chatbots based on their own course material, but also learn the fundamentals of AI through the model creation process. The primary concern of this paper is the creation of an interface for students to learn the principles of artificial intelligence by using a natural language pipeline to train a customized model to answer questions based on their own school curriculums. The model uses contexts given by their instructor, such as chapters of a textbook, to answer questions and is deployed on an interactive chatbot/voice agent. The pipeline teaches students data collection, data augmentation, intent recognition, and question answering by having them work through each of these processes while creating their AI agent, diverging from previous chatbot work where students and teachers use the bots as black-boxes with no abilities for customization or the bots lack AI capabilities, with the majority of dialogue scripts being rule-based. In addition, our tool is designed to make each step of this pipeline intuitive for students at a middle-school level. Further work primarily lies in providing our tool to schools and seeking student and teacher evaluations.

Non-invasive prostate cancer detection from MRI has the potential to revolutionize patient care by providing early detection of clinically-significant disease (ISUP grade group >= 2), but has thus far shown limited positive predictive value. To address this, we present an MRI-based deep learning method for predicting clinically significant prostate cancer applicable to a patient population with subsequent ground truth biopsy results ranging from benign pathology to ISUP grade group~5. Specifically, we demonstrate that mixed supervision via diverse histopathological ground truth improves classification performance despite the cost of reduced concordance with image-based segmentation. That is, where prior approaches have utilized pathology results as ground truth derived from targeted biopsies and whole-mount prostatectomy to strongly supervise the localization of clinically significant cancer, our approach also utilizes weak supervision signals extracted from nontargeted systematic biopsies with regional localization to improve overall performance. Our key innovation is performing regression by distribution rather than simply by value, enabling use of additional pathology findings traditionally ignored by deep learning strategies. We evaluated our model on a dataset of 973 (testing n=160) multi-parametric prostate MRI exams collected at UCSF from 2015-2018 followed by MRI/ultrasound fusion (targeted) biopsy and systematic (nontargeted) biopsy of the prostate gland, demonstrating that deep networks trained with mixed supervision of histopathology can significantly exceed the performance of the Prostate Imaging-Reporting and Data System (PI-RADS) clinical standard for prostate MRI interpretation.

Transfer operators offer linear representations and global, physically meaningful features of nonlinear dynamical systems. Discovering transfer operators, such as the Koopman operator, require careful crafted dictionaries of observables, acting on states of the dynamical system. This is ad hoc and requires the full dataset for evaluation. In this paper, we offer an optimization scheme to allow joint learning of the observables and Koopman operator with online data. Our results show we are able to reconstruct the evolution and represent the global features of complex dynamical systems.

Modern deep neural networks tend to be evaluated on static test sets. One shortcoming of this is the fact that these deep neural networks cannot be easily evaluated for robustness issues with respect to specific scene variations. For example, it is hard to study the robustness of these networks to variations of object scale, object pose, scene lighting and 3D occlusions. The main reason is that collecting real datasets with fine-grained naturalistic variations of sufficient scale can be extremely time-consuming and expensive. In this work, we present Counterfactual Simulation Testing, a counterfactual framework that allows us to study the robustness of neural networks with respect to some of these naturalistic variations by building realistic synthetic scenes that allow us to ask counterfactual questions to the models, ultimately providing answers to questions such as "Would your classification still be correct if the object were viewed from the top?" or "Would your classification still be correct if the object were partially occluded by another object?". Our method allows for a fair comparison of the robustness of recently released, state-of-the-art Convolutional Neural Networks and Vision Transformers, with respect to these naturalistic variations. We find evidence that ConvNext is more robust to pose and scale variations than Swin, that ConvNext generalizes better to our simulated domain and that Swin handles partial occlusion better than ConvNext. We also find that robustness for all networks improves with network scale and with data scale and variety. We release the Naturalistic Variation Object Dataset (NVD), a large simulated dataset of 272k images of everyday objects with naturalistic variations such as object pose, scale, viewpoint, lighting and occlusions. Project page: https://counterfactualsimulation.github.io

While transformers have greatly boosted performance in semantic segmentation, domain adaptive transformers are not yet well explored. We identify that the domain gap can cause discrepancies in self-attention. Due to this gap, the transformer attends to spurious regions or pixels, which deteriorates accuracy on the target domain. We propose to perform adaptation on attention maps with cross-domain attention layers that share features between the source and the target domains. Specifically, we impose consistency between predictions from cross-domain attention and self-attention modules to encourage similar distribution in the attention and output of the model across domains, i.e., attention-level and output-level alignment. We also enforce consistency in attention maps between different augmented views to further strengthen the attention-based alignment. Combining these two components, our method mitigates the discrepancy in attention maps across domains and further boosts the performance of the transformer under unsupervised domain adaptation settings. Our model outperforms the existing state-of-the-art baseline model on three widely used benchmarks, including GTAV-to-Cityscapes by 1.3 percent point (pp), Synthia-to-Cityscapes by 0.6 pp, and Cityscapes-to-ACDC by 1.1 pp, on average. Additionally, we verify the effectiveness and generalizability of our method through extensive experiments. Our code will be publicly available.

Physics-informed neural networks have been widely applied to learn general parametric solutions of differential equations. Here, we propose a neural network to discover parametric eigenvalue and eigenfunction surfaces of quantum systems. We apply our method to solve the hydrogen molecular ion. This is an ab-initio deep learning method that solves the Schrodinger equation with the Coulomb potential yielding realistic wavefunctions that include a cusp at the ion positions. The neural solutions are continuous and differentiable functions of the interatomic distance and their derivatives are analytically calculated by applying automatic differentiation. Such a parametric and analytical form of the solutions is useful for further calculations such as the determination of force fields.