In recent years, we have witnessed tremendous progress in deep reinforcement learning(RL)fortaskssuchasGo,Chess,videogames,androbotcontrol.

Prior work has found that pretrained language models (LMs) fine-tuned with different random seeds can achieve similar in-domain performance but generalize differently on tests of syntactic generalization. In this work, we show that, even within a single model, we can find multiple subnetworks that perform similarly in-domain, but generalize vastly differently. To better understand these phenomena, we investigate if they can be understood in terms of "competing subnetworks": the model initially represents a variety of distinct algorithms, corresponding to different subnetworks, and generalization occurs when it ultimately converges to one. This explanation has been used to account for generalization in simple algorithmic tasks ("grokking"). Instead of finding competing subnetworks, we find that all subnetworks -- whether they generalize or not -- share a set of attention heads, which we refer to as the heuristic core. Further analysis suggests that these attention heads emerge early in training and compute shallow, non-generalizing features. The model learns to generalize by incorporating additional attention heads, which depend on the outputs of the "heuristic" heads to compute higher-level features. Overall, our results offer a more detailed picture of the mechanisms for syntactic generalization in pretrained LMs.

A traditional assurance case employs a positive argument in which reasoning steps, grounded on evidence and assumptions, sustain a top claim that has external significance. Human judgement is required to check the evidence, the assumptions, and the narrative justifications for the reasoning steps; if all are assessed good, then the top claim can be accepted. A valid concern about this process is that human judgement is fallible and prone to confirmation bias. The best defense against this concern is vigorous and skeptical debate and discussion in the manner of a dialectic or Socratic dialog. There is merit in recording aspects of this discussion for the benefit of subsequent developers and assessors. Defeaters are a means doing this: they express doubts about aspects of the argument and can be developed into subcases that confirm or refute the doubts, and can record them as documentation to assist future consideration. This report describes how defeaters, and multiple levels of defeaters, should be represented and assessed in Assurance 2.0 and its Clarissa/ASCE tool support. These mechanisms also support eliminative argumentation, which is a contrary approach to assurance, favored by some, that uses a negative argument to refute all reasons why the top claim could be false.

Optimal experimental design (OED) provides a systematic approach to quantify and maximize the value of experimental data. Under a Bayesian approach, conventional OED maximizes the expected information gain (EIG) on model parameters. However, we are often interested in not the parameters themselves, but predictive quantities of interest (QoIs) that depend on the parameters in a nonlinear manner. We present a computational framework of predictive goal-oriented OED (GO-OED) suitable for nonlinear observation and prediction models, which seeks the experimental design providing the greatest EIG on the QoIs. In particular, we propose a nested Monte Carlo estimator for the QoI EIG, featuring Markov chain Monte Carlo for posterior sampling and kernel density estimation for evaluating the posterior-predictive density and its Kullback-Leibler divergence from the prior-predictive. The GO-OED design is then found by maximizing the EIG over the design space using Bayesian optimization. We demonstrate the effectiveness of the overall nonlinear GO-OED method, and illustrate its differences versus conventional non-GO-OED, through various test problems and an application of sensor placement for source inversion in a convection-diffusion field.

In this paper, we study the two-level ski-rental problem,where a user needs to fulfill a sequence of demands for multiple items by choosing one of the three payment options: paying for the on-demand usage (i.e., rent), buying individual items (i.e., single purchase), and buying all the items (i.e., combo purchase). Without knowing future demands, the user aims to minimize the total cost (i.e., the sum of the rental, single purchase, and combo purchase costs) by balancing the trade-off between the expensive upfront costs (for purchase) and the potential future expenses (for rent). We first design a robust online algorithm (RDTSR) that offers a worst-case performance guarantee. While online algorithms are robust against the worst-case scenarios, they are often overly cautious and thus suffer a poor average performance in typical scenarios. On the other hand, Machine Learning (ML) algorithms typically show promising average performance in various applications but lack worst-case performance guarantees. To harness the benefits of both methods, we develop a learning-augmented algorithm (LADTSR) by integrating ML predictions into the robust online algorithm, which outperforms the robust online algorithm under accurate predictions while ensuring worst-case performance guarantees even when predictions are inaccurate. Finally, we conduct numerical experiments on both synthetic and real-world trace data to corroborate the effectiveness of our approach.

Prediction models are popular in medical research and practice. By predicting an outcome of interest for specific patients, these models may help inform difficult treatment decisions, and are often hailed as the poster children for personalized, data-driven healthcare. We show however, that using prediction models for decision making can lead to harmful decisions, even when the predictions exhibit good discrimination after deployment. These models are harmful self-fulfilling prophecies: their deployment harms a group of patients but the worse outcome of these patients does not invalidate the predictive power of the model. Our main result is a formal characterization of a set of such prediction models. Next we show that models that are well calibrated before and after deployment are useless for decision making as they made no change in the data distribution. These results point to the need to revise standard practices for validation, deployment and evaluation of prediction models that are used in medical decisions.

Corruption is frequently observed in collected data and has been extensively studied in machine learning under different corruption models. Despite this, there remains a limited understanding of how these models relate such that a unified view of corruptions and their consequences on learning is still lacking. In this work, we formally analyze corruption models at the distribution level through a general, exhaustive framework based on Markov kernels. We highlight the existence of intricate joint and dependent corruptions on both labels and attributes, which are rarely touched by existing research. Further, we show how these corruptions affect standard supervised learning by analyzing the resulting changes in Bayes Risk. Our findings offer qualitative insights into the consequences of "more complex" corruptions on the learning problem, and provide a foundation for future quantitative comparisons. Applications of the framework include corruption-corrected learning, a subcase of which we study in this paper by theoretically analyzing loss correction with respect to different corruption instances.

An assurance case is intended to provide justifiable confidence in the truth of its top claim, which typically concerns safety or security. A natural question is then "how much" confidence does the case provide? We argue that confidence cannot be reduced to a single attribute or measurement. Instead, we suggest it should be based on attributes that draw on three different perspectives: positive, negative, and residual doubts. Positive Perspectives consider the extent to which the evidence and overall argument of the case combine to make a positive statement justifying belief in its claims. We set a high bar for justification, requiring it to be indefeasible. The primary positive measure for this is soundness, which interprets the argument as a logical proof. Confidence in evidence can be expressed probabilistically and we use confirmation measures to ensure that the "weight" of evidence crosses some threshold. In addition, probabilities can be aggregated from evidence through the steps of the argument using probability logics to yield what we call probabilistic valuations for the claims. Negative Perspectives record doubts and challenges to the case, typically expressed as defeaters, and their exploration and resolution. Assurance developers must guard against confirmation bias and should vigorously explore potential defeaters as they develop the case, and should record them and their resolution to avoid rework and to aid reviewers. Residual Doubts: the world is uncertain so not all potential defeaters can be resolved. We explore risks and may deem them acceptable or unavoidable. It is crucial however that these judgments are conscious ones and that they are recorded in the assurance case. This report examines the perspectives in detail and indicates how Clarissa, our prototype toolset for Assurance 2.0, assists in their evaluation.