关于自适应梯度方法等自适应梯度方法等训练动力的知之甚少。在本文中，我们阐明了这些算法在全批处理和足够大的批处理设置中的行为。具体而言，我们从经验上证明，在全批训练中，预处理的Hessian的最大特征值通常在某个数值下平衡 - 梯度下降算法的稳定性阈值。对于带有步长$ \ eta $和$ \ beta_1 = 0.9 $的Adam，此稳定性阈值为$ 38/\ eta $。在Minibatch培训期间发生了类似的影响，尤其是随着批处理大小的增长。然而，即使自适应方法在``稳定性的自适应边缘''（AEOS）上训练，但它们在该制度中的行为与EOS的非自适应方法的行为有很大不同。 EOS处的非自适应算法被阻止进入损失景观的高曲率区域，而AEOS的自适应梯度方法可以继续前进到高外观区域，同时适应预先调节器以补偿。我们的发现可以成为社区对深度学习中适应性梯度方法的未来理解的基础。

人工智能的最新趋势是将验证的模型用于语言和视觉任务，这些模型已经实现了非凡的表现，但也令人困惑。因此，以各种方式探索这些模型的能力对该领域至关重要。在本文中，我们探讨了模型的可靠性，在其中我们将可靠的模型定义为一个不仅可以实现强大的预测性能，而且在许多涉及不确定性（例如选择性预测，开放式设置识别）的决策任务上，在许多决策任务上表现出色，而且表现良好。强大的概括（例如，准确性和适当的评分规则，例如在分布数据集中和分发数据集上的对数可能性）和适应性（例如，主动学习，几乎没有射击不确定性）。我们设计了40个数据集的10种任务类型，以评估视觉和语言域上可靠性的不同方面。为了提高可靠性，我们分别开发了VIT-PLEX和T5-PLEX，分别针对视觉和语言方式扩展了大型模型。 PLEX极大地改善了跨可靠性任务的最先进，并简化了传统协议，因为它可以改善开箱即用的性能，并且不需要设计分数或为每个任务调整模型。我们演示了高达1B参数的模型尺寸的缩放效果，并预处理数据集大小最多4B示例。我们还展示了PLEX在具有挑战性的任务上的功能，包括零射门的开放式识别，主动学习和对话语言理解中的不确定性。

贝叶斯优化（BO）已成为许多昂贵现实世界功能的全球优化的流行策略。与普遍认为BO适合优化黑框功能的信念相反，它实际上需要有关这些功能特征的域知识才能成功部署BO。这样的领域知识通常表现在高斯流程先验中，这些先验指定了有关功能的初始信念。但是，即使有专家知识，选择先验也不是一件容易的事。对于复杂的机器学习模型上的超参数调谐问题尤其如此，在这种模型中，调整目标的景观通常很难理解。我们寻求一种设定这些功能性先验的替代实践。特别是，我们考虑了从类似功能的数据中，使我们可以先验地进行更紧密的分布。为了验证我们在现实的模型培训设置中的方法，我们通过训练在流行图像和文本数据集上的数以万计的近状态模型配置来收集了大型多任务超参数调谐数据集，以及蛋白质序列数据集。我们的结果表明，平均而言，我们的方法能够比最佳竞争方法更有效地定位良好的超参数。

贝叶斯优化（BO）已成为许多昂贵现实世界功能的全球优化的流行策略。与普遍认为BO适合优化黑框功能的信念相反，它实际上需要有关这些功能特征的域知识才能成功部署BO。这样的领域知识通常表现在高斯流程先验中，这些先验指定了有关功能的初始信念。但是，即使有专家知识，选择先验也不是一件容易的事。对于复杂的机器学习模型上的超参数调谐问题尤其如此，在这种模型中，调整目标的景观通常很难理解。我们寻求一种设定这些功能性先验的替代实践。特别是，我们考虑了从类似功能的数据中，使我们可以先验地进行更紧密的分布。从理论上讲，我们与预先训练的先验表示对BO的遗憾。为了验证我们在现实的模型培训设置中的方法，我们通过训练在流行图像和文本数据集上的数以万计的近状态模型配置来收集了大型多任务超参数调谐数据集，以及蛋白质序列数据集。我们的结果表明，平均而言，我们的方法能够比最佳竞争方法更有效地定位良好的超参数。

对不确定度和鲁棒性的高质量估计对于众多现实世界的应用来说至关重要，特别是对于深入学习，这是利用许多部署的ML系统。因此，比较改善这些估计的技术的能力对于研究和实践相似非常重要。然而，由于一系列原因，通常缺乏方法的竞争比较，包括：计算广泛调整的可用性，加入足够多的基线，以及用于再现性的具体文件。在本文中，我们介绍了不确定性的基线：在各种任务中的标准和最先进的深度学习方法的高质量实现。从本撰写中，集合跨越9项方法，每个方法都有至少5个度量。每个基线都是一个独立的实验管道，易于可重复使用和可伸缩的部件。我们的目标是提供具有新方法或应用的实验的即时出发点。此外，我们还提供模型检查点，实验输出为Python笔记本，以及用于比较结果的排行榜。代码在https://github.com/google/uncertainty-baselines。

Modern machine learning methods including deep learning have achieved great success in predictive accuracy for supervised learning tasks, but may still fall short in giving useful estimates of their predictive uncertainty. Quantifying uncertainty is especially critical in real-world settings, which often involve input distributions that are shifted from the training distribution due to a variety of factors including sample bias and non-stationarity. In such settings, well calibrated uncertainty estimates convey information about when a model's output should (or should not) be trusted. Many probabilistic deep learning methods, including Bayesian-and non-Bayesian methods, have been proposed in the literature for quantifying predictive uncertainty, but to our knowledge there has not previously been a rigorous largescale empirical comparison of these methods under dataset shift. We present a largescale benchmark of existing state-of-the-art methods on classification problems and investigate the effect of dataset shift on accuracy and calibration. We find that traditional post-hoc calibration does indeed fall short, as do several other previous methods. However, some methods that marginalize over models give surprisingly strong results across a broad spectrum of tasks.

Course load analytics (CLA) inferred from LMS and enrollment features can offer a more accurate representation of course workload to students than credit hours and potentially aid in their course selection decisions. In this study, we produce and evaluate the first machine-learned predictions of student course load ratings and generalize our model to the full 10,000 course catalog of a large public university. We then retrospectively analyze longitudinal differences in the semester load of student course selections throughout their degree. CLA by semester shows that a student's first semester at the university is among their highest load semesters, as opposed to a credit hour-based analysis, which would indicate it is among their lowest. Investigating what role predicted course load may play in program retention, we find that students who maintain a semester load that is low as measured by credit hours but high as measured by CLA are more likely to leave their program of study. This discrepancy in course load is particularly pertinent in STEM and associated with high prerequisite courses. Our findings have implications for academic advising, institutional handling of the freshman experience, and student-facing analytics to help students better plan, anticipate, and prepare for their selected courses.

We present a differentiable formulation of rigid-body contact dynamics for objects and robots represented as compositions of convex primitives. Existing optimization-based approaches simulating contact between convex primitives rely on a bilevel formulation that separates collision detection and contact simulation. These approaches are unreliable in realistic contact simulation scenarios because isolating the collision detection problem introduces contact location non-uniqueness. Our approach combines contact simulation and collision detection into a unified single-level optimization problem. This disambiguates the collision detection problem in a physics-informed manner. Compared to previous differentiable simulation approaches, our formulation features improved simulation robustness and a reduction in computational complexity by more than an order of magnitude. We illustrate the contact and collision differentiability on a robotic manipulation task requiring optimization-through-contact. We provide a numerically efficient implementation of our formulation in the Julia language called Silico.jl.

As Artificial and Robotic Systems are increasingly deployed and relied upon for real-world applications, it is important that they exhibit the ability to continually learn and adapt in dynamically-changing environments, becoming Lifelong Learning Machines. Continual/lifelong learning (LL) involves minimizing catastrophic forgetting of old tasks while maximizing a model's capability to learn new tasks. This paper addresses the challenging lifelong reinforcement learning (L2RL) setting. Pushing the state-of-the-art forward in L2RL and making L2RL useful for practical applications requires more than developing individual L2RL algorithms; it requires making progress at the systems-level, especially research into the non-trivial problem of how to integrate multiple L2RL algorithms into a common framework. In this paper, we introduce the Lifelong Reinforcement Learning Components Framework (L2RLCF), which standardizes L2RL systems and assimilates different continual learning components (each addressing different aspects of the lifelong learning problem) into a unified system. As an instantiation of L2RLCF, we develop a standard API allowing easy integration of novel lifelong learning components. We describe a case study that demonstrates how multiple independently-developed LL components can be integrated into a single realized system. We also introduce an evaluation environment in order to measure the effect of combining various system components. Our evaluation environment employs different LL scenarios (sequences of tasks) consisting of Starcraft-2 minigames and allows for the fair, comprehensive, and quantitative comparison of different combinations of components within a challenging common evaluation environment.

Rearrangement puzzles are variations of rearrangement problems in which the elements of a problem are potentially logically linked together. To efficiently solve such puzzles, we develop a motion planning approach based on a new state space that is logically factored, integrating the capabilities of the robot through factors of simultaneously manipulatable joints of an object. Based on this factored state space, we propose less-actions RRT (LA-RRT), a planner which optimizes for a low number of actions to solve a puzzle. At the core of our approach lies a new path defragmentation method, which rearranges and optimizes consecutive edges to minimize action cost. We solve six rearrangement scenarios with a Fetch robot, involving planar table puzzles and an escape room scenario. LA-RRT significantly outperforms the next best asymptotically-optimal planner by 4.01 to 6.58 times improvement in final action cost.