The deployment of machine learning models in safety-critical applications comes with the expectation that such models will perform well over a range of contexts (e.g., a vision model for classifying street signs should work in rural, city, and highway settings under varying lighting/weather conditions). However, these one-size-fits-all models are typically optimized for average case performance, encouraging them to achieve high performance in nominal conditions but exposing them to unexpected behavior in challenging or rare contexts. To address this concern, we develop a new method for training context-dependent models. We extend Bridge-Mode Connectivity (BMC) (Garipov et al., 2018) to train an infinite ensemble of models over a continuous measure of context such that we can sample model parameters specifically tuned to the corresponding evaluation context. We explore the definition of context in image classification tasks through multiple lenses including changes in the risk profile, long-tail image statistics/appearance, and context-dependent distribution shift. We develop novel extensions of the BMC optimization for each of these cases and our experiments demonstrate that model performance can be successfully tuned to context in each scenario.

用于计算机视觉任务的深度神经网络在越来越安全 - 严重和社会影响的应用中部署，激励需要在各种，天然存在的成像条件下关闭模型性能的差距。在包括对抗机器学习的多种上下文中尤为色难地使用的鲁棒性，然后指在自然诱导的图像损坏或改变下保持模型性能。我们进行系统审查，以识别，分析和总结当前定义以及对计算机愿景深度学习中的非对抗鲁棒性的进展。我们发现，该研究领域已经收到了相对于对抗机器学习的不成比例地注意力，但存在显着的稳健性差距，这些差距通常表现在性能下降中与对抗条件相似。为了在上下文中提供更透明的稳健性定义，我们引入了数据生成过程的结构因果模型，并将非对抗性鲁棒性解释为模型在损坏的图像上的行为，其对应于来自未纳入数据分布的低概率样本。然后，我们确定提高神经网络鲁棒性的关键架构，数据增强和优化策略。这种稳健性的这种因果观察表明，目前文献中的常见做法，关于鲁棒性策略和评估，对应于因果概念，例如软干预导致成像条件的决定性分布。通过我们的调查结果和分析，我们提供了对未来研究如何可能介意这种明显和显着的非对抗的鲁棒性差距的观点。

虽然对抗性攻击检测得到了相当大的关注，但它仍然是两个观点的基本上具有挑战性的问题。首先，虽然威胁模型可以明确定义，但攻击者策略可能在这些限制范围内仍然很大。因此，检测应被视为开放式问题，与大多数电流检测方法相比，站立相反。这些方法采用封闭式视图和火车二进制探测器，从而偏置检测探测器训练期间看到的攻击。其次，有限的信息可在测试时间上获得，并且通常通过滋扰因子混淆，包括标签和图像的底层内容。我们通过基于随机子空间分析的新策略来解决这些挑战。我们提出了一种利用随机投影的性质的技术，以表征在各种子空间中的清洁和对抗性示例的行为。模型激活的自我一致性（或不一致）被利用从对抗例中辨别清洁。性能评估表明，我们的技术（$ AUC \在[0.92,0.98] $）优于竞争检测策略（$ AUC \在[0.30,0.79]中），同时仍然真正无法对攻击战略（针对目标/未确定的攻击） ）。它还需要显着更少的校准数据（仅由干净的例子组成）而不是实现这种性能的竞争方法。

Feature selection helps reduce data acquisition costs in ML, but the standard approach is to train models with static feature subsets. Here, we consider the dynamic feature selection (DFS) problem where a model sequentially queries features based on the presently available information. DFS is often addressed with reinforcement learning (RL), but we explore a simpler approach of greedily selecting features based on their conditional mutual information. This method is theoretically appealing but requires oracle access to the data distribution, so we develop a learning approach based on amortized optimization. The proposed method is shown to recover the greedy policy when trained to optimality and outperforms numerous existing feature selection methods in our experiments, thus validating it as a simple but powerful approach for this problem.

We demonstrate how efficient autonomous drone swarms can be in detecting and tracking occluded targets in densely forested areas, such as lost people during search and rescue missions. Exploration and optimization of local viewing conditions, such as occlusion density and target view obliqueness, provide much faster and much more reliable results than previous, blind sampling strategies that are based on pre-defined waypoints. An adapted real-time particle swarm optimization and a new objective function are presented that are able to deal with dynamic and highly random through-foliage conditions. Synthetic aperture sensing is our fundamental sampling principle, and drone swarms are employed to approximate the optical signals of extremely wide and adaptable airborne lenses.

Sequential testing, always-valid $p$-values, and confidence sequences promise flexible statistical inference and on-the-fly decision making. However, unlike fixed-$n$ inference based on asymptotic normality, existing sequential tests either make parametric assumptions and end up under-covering/over-rejecting when these fail or use non-parametric but conservative concentration inequalities and end up over-covering/under-rejecting. To circumvent these issues, we sidestep exact at-least-$\alpha$ coverage and focus on asymptotically exact coverage and asymptotic optimality. That is, we seek sequential tests whose probability of ever rejecting a true hypothesis asymptotically approaches $\alpha$ and whose expected time to reject a false hypothesis approaches a lower bound on all tests with asymptotic coverage at least $\alpha$, both under an appropriate asymptotic regime. We permit observations to be both non-parametric and dependent and focus on testing whether the observations form a martingale difference sequence. We propose the universal sequential probability ratio test (uSPRT), a slight modification to the normal-mixture sequential probability ratio test, where we add a burn-in period and adjust thresholds accordingly. We show that even in this very general setting, the uSPRT is asymptotically optimal under mild generic conditions. We apply the results to stabilized estimating equations to test means, treatment effects, etc. Our results also provide corresponding guarantees for the implied confidence sequences. Numerical simulations verify our guarantees and the benefits of the uSPRT over alternatives.

Large language models (LLMs) have demonstrated impressive capabilities in natural language understanding and generation, but the quality bar for medical and clinical applications is high. Today, attempts to assess models' clinical knowledge typically rely on automated evaluations on limited benchmarks. There is no standard to evaluate model predictions and reasoning across a breadth of tasks. To address this, we present MultiMedQA, a benchmark combining six existing open question answering datasets spanning professional medical exams, research, and consumer queries; and HealthSearchQA, a new free-response dataset of medical questions searched online. We propose a framework for human evaluation of model answers along multiple axes including factuality, precision, possible harm, and bias. In addition, we evaluate PaLM (a 540-billion parameter LLM) and its instruction-tuned variant, Flan-PaLM, on MultiMedQA. Using a combination of prompting strategies, Flan-PaLM achieves state-of-the-art accuracy on every MultiMedQA multiple-choice dataset (MedQA, MedMCQA, PubMedQA, MMLU clinical topics), including 67.6% accuracy on MedQA (US Medical License Exam questions), surpassing prior state-of-the-art by over 17%. However, human evaluation reveals key gaps in Flan-PaLM responses. To resolve this we introduce instruction prompt tuning, a parameter-efficient approach for aligning LLMs to new domains using a few exemplars. The resulting model, Med-PaLM, performs encouragingly, but remains inferior to clinicians. We show that comprehension, recall of knowledge, and medical reasoning improve with model scale and instruction prompt tuning, suggesting the potential utility of LLMs in medicine. Our human evaluations reveal important limitations of today's models, reinforcing the importance of both evaluation frameworks and method development in creating safe, helpful LLM models for clinical applications.

Transformers have been essential to pretraining success in NLP. Other architectures have been used, but require attention layers to match benchmark accuracy. This work explores pretraining without attention. We test recently developed routing layers based on state-space models (SSM) and model architectures based on multiplicative gating. Used together these modeling choices have a large impact on pretraining accuracy. Empirically the proposed Bidirectional Gated SSM (BiGS) replicates BERT pretraining results without attention and can be extended to long-form pretraining of 4096 tokens without approximation.

In this paper, we present strong baselines for the task of Feedback Comment Generation for Writing Learning. Given a sentence and an error span, the task is to generate a feedback comment explaining the error. Sentences and feedback comments are both in English. We experiment with LLMs and also create multiple pseudo datasets for the task, investigating how it affects the performance of our system. We present our results for the task along with extensive analysis of the generated comments with the aim of aiding future studies in feedback comment generation for English language learners.

In order for automated mobile vehicles to navigate in the real world with minimal collision risks, it is necessary for their planning algorithms to consider uncertainties from measurements and environmental disturbances. In this paper, we consider analytical solutions for a conservative approximation of the mutual probability of collision between two robotic vehicles in the presence of such uncertainties. Therein, we present two methods, which we call unitary scaling and principal axes rotation, for decoupling the bivariate integral required for efficient approximation of the probability of collision between two vehicles including orientation effects. We compare the conservatism of these methods analytically and numerically. By closing a control loop through a model predictive guidance scheme, we observe through Monte-Carlo simulations that directly implementing collision avoidance constraints from the conservative approximations remains infeasible for real-time planning. We then propose and implement a convexification approach based on the tightened collision constraints that significantly improves the computational efficiency and robustness of the predictive guidance scheme.