Pandora's Box is a fundamental stochastic optimization problem, where the decision-maker must find a good alternative while minimizing the search cost of exploring the value of each alternative. In the original formulation, it is assumed that accurate distributions are given for the values of all the alternatives, while recent work studies the online variant of Pandora's Box where the distributions are originally unknown. In this work, we study Pandora's Box in the online setting, while incorporating context. At every round, we are presented with a number of alternatives each having a context, an exploration cost and an unknown value drawn from an unknown distribution that may change at every round. Our main result is a no-regret algorithm that performs comparably well to the optimal algorithm which knows all prior distributions exactly. Our algorithm works even in the bandit setting where the algorithm never learns the values of the alternatives that were not explored. The key technique that enables our result is a novel modification of the realizability condition in contextual bandits that connects a context to a sufficient statistic of each alternative's distribution (its "reservation value") rather than its mean.

Models that can predict the occurrence of events ahead of time with low false-alarm rates are critical to the acceptance of decision support systems in the medical community. This challenging task is typically treated as a simple binary classification, ignoring temporal dependencies between samples, whereas we propose to exploit this structure. We first introduce a common theoretical framework unifying dynamic survival analysis and early event prediction. Following an analysis of objectives from both fields, we propose Temporal Label Smoothing (TLS), a simpler, yet best-performing method that preserves prediction monotonicity over time. By focusing the objective on areas with a stronger predictive signal, TLS improves performance over all baselines on two large-scale benchmark tasks. Gains are particularly notable along clinically relevant measures, such as event recall at low false-alarm rates. TLS reduces the number of missed events by up to a factor of two over previously used approaches in early event prediction.

Detecting out-of-distribution (OOD) inputs is crucial for the safe deployment of natural language processing (NLP) models. Though existing methods, especially those based on the statistics in the feature space of fine-tuned pre-trained language models (PLMs), are claimed to be effective, their effectiveness on different types of distribution shifts remains underexplored. In this work, we take the first step to comprehensively evaluate the mainstream textual OOD detection methods for detecting semantic and non-semantic shifts. We find that: (1) no existing method behaves well in both settings; (2) fine-tuning PLMs on in-distribution data benefits detecting semantic shifts but severely deteriorates detecting non-semantic shifts, which can be attributed to the distortion of task-agnostic features. To alleviate the issue, we present a simple yet effective general OOD score named GNOME that integrates the confidence scores derived from the task-agnostic and task-specific representations. Experiments show that GNOME works well in both semantic and non-semantic shift scenarios, and further brings significant improvement on two cross-task benchmarks where both kinds of shifts simultaneously take place. Our code is available at https://github.com/lancopku/GNOME.

Curriculum Data Augmentation (CDA) improves neural models by presenting synthetic data with increasing difficulties from easy to hard. However, traditional CDA simply treats the ratio of word perturbation as the difficulty measure and goes through the curriculums only once. This paper presents \textbf{PCC}: \textbf{P}araphrasing with Bottom-k Sampling and \textbf{C}yclic Learning for \textbf{C}urriculum Data Augmentation, a novel CDA framework via paraphrasing, which exploits the textual paraphrase similarity as the curriculum difficulty measure. We propose a curriculum-aware paraphrase generation module composed of three units: a paraphrase candidate generator with bottom-k sampling, a filtering mechanism and a difficulty measure. We also propose a cyclic learning strategy that passes through the curriculums multiple times. The bottom-k sampling is proposed to generate super-hard instances for the later curriculums. Experimental results on few-shot text classification as well as dialogue generation indicate that PCC surpasses competitive baselines. Human evaluation and extensive case studies indicate that bottom-k sampling effectively generates super-hard instances, and PCC significantly improves the baseline dialogue agent.

We study the task of prompting large-scale language models to perform multistep reasoning. Existing work shows that when prompted with a chain of thoughts (CoT), sequences of short sentences describing intermediate reasoning steps towards a final answer, large language models can generate new reasoning chains and predict answers for new inputs. A central question is which reasoning examples make the most effective prompts. In this work, we propose complexitybased prompting, a simple and effective example selection scheme for multi-step reasoning. We show that prompts with higher reasoning complexity, i.e., chains with more reasoning steps, achieve substantially better performance on multistep reasoning tasks over strong baselines. We further extend our complexitybased criteria from prompting (selecting inputs) to decoding (selecting outputs), where we sample multiple reasoning chains from the model, then choose the majority of generated answers from complex reasoning chains (over simple chains). When used to prompt GPT-3 and Codex, our approach substantially improves multi-step reasoning accuracy and achieves new state-of-the-art (SOTA) performance on three math benchmarks (GSM8K, MultiArith, and MathQA) and two BigBenchHard tasks (Date Understanding and Penguins), with an average +5.3 and up to +18 accuracy improvements. Compared with existing example selection schemes like manual tuning or retrieval-based selection, selection based on reasoning complexity is intuitive, easy to implement, and annotation-efficient. Further results demonstrate the robustness of performance gains from complex prompts under format perturbation and distribution shift. We consider the problem of prompting large language models for multi-step reasoning. Recent breakthroughs (Wei et al., 2022b; Wang et al., 2022b) show that language models, when large enough (>100B parameters), exhibit the emergent ability (Wei et al., 2022a) of performing complex multi-step reasoning when provided with only a few reasoning examples.

Few-shot learning allows pre-trained language models to adapt to downstream tasks while using a limited number of training examples. However, practical applications are limited when all model parameters must be optimized. In this work we apply a new technique for parameter efficient few shot learning while adopting a strict definition of parameter efficiency. Our training method combines 1) intermediate training by reformulating natural language tasks as entailment tasks \cite{wang_entailment_2021} and 2) differentiable optimization of template and label tokens \cite{zhang_differentiable_2021}. We quantify the tradeoff between parameter efficiency and performance in the few-shot regime and propose a simple model agnostic approach that can be extended to any task By achieving competitive performance while only optimizing 3\% of a model's parameters and allowing for batched inference, we allow for more efficient practical deployment of models.

Many instances of similar or almost-identical industrial machines or tools are often deployed at once, or in quick succession. For instance, a particular model of air compressor may be installed at hundreds of customers. Because these tools perform distinct but highly similar tasks, it is interesting to be able to quickly produce a high-quality controller for machine $N+1$ given the controllers already produced for machines $1..N$. This is even more important when the controllers are learned through Reinforcement Learning, as training takes time, energy and other resources. In this paper, we apply Policy Intersection, a Policy Shaping method, to help a Reinforcement Learning agent learn to solve a new variant of a compressors control problem faster, by transferring knowledge from several previously learned controllers. We show that our approach outperforms loading an old controller, and significantly improves performance in the long run.

Recent deep learning (DL) applications are mostly built on top of DL libraries. The quality assurance of these libraries is critical to the dependable deployment of DL applications. Techniques have been proposed to generate various DL models and apply them to test these libraries. However, their test effectiveness is constrained by the diversity of layer API calls in their generated DL models. Our study reveals that these techniques can cover at most 34.1% layer inputs, 25.9% layer parameter values, and 15.6% layer sequences. As a result, we find that many bugs arising from specific layer API calls (i.e., specific layer inputs, parameter values, or layer sequences) can be missed by existing techniques. Because of this limitation, we propose COMET to effectively generate DL models with diverse layer API calls for DL library testing. COMET: (1) designs a set of mutation operators and a coverage-based search algorithm to diversify layer inputs, layer parameter values, and layer sequences in DL models. (2) proposes a model synthesis method to boost the test efficiency without compromising the layer API call diversity. Our evaluation result shows that COMET outperforms baselines by covering twice as many layer inputs (69.7% vs. 34.1%), layer parameter values (50.2% vs. 25.9%), and layer sequences (39.0% vs. 15.6%) as those by the state-of-the-art. Moreover, COMET covers 3.4% more library branches than those by existing techniques. Finally, COMET detects 32 new bugs in the latest version of eight popular DL libraries, including TensorFlow and MXNet, with 21 of them confirmed by DL library developers and 7 of those confirmed bugs have been fixed by developers.

Self-distillation (SD) is the process of first training a \enquote{teacher} model and then using its predictions to train a \enquote{student} model with the \textit{same} architecture. Specifically, the student's objective function is $\big(\xi*\ell(\text{teacher's predictions}, \text{ student's predictions}) + (1-\xi)*\ell(\text{given labels}, \text{ student's predictions})\big)$, where $\ell$ is some loss function and $\xi$ is some parameter $\in [0,1]$. Empirically, SD has been observed to provide performance gains in several settings. In this paper, we theoretically characterize the effect of SD in two supervised learning problems with \textit{noisy labels}. We first analyze SD for regularized linear regression and show that in the high label noise regime, the optimal value of $\xi$ that minimizes the expected error in estimating the ground truth parameter is surprisingly greater than 1. Empirically, we show that $\xi > 1$ works better than $\xi \leq 1$ even with the cross-entropy loss for several classification datasets when 50\% or 30\% of the labels are corrupted. Further, we quantify when optimal SD is better than optimal regularization. Next, we analyze SD in the case of logistic regression for binary classification with random label corruption and quantify the range of label corruption in which the student outperforms the teacher in terms of accuracy. To our knowledge, this is the first result of its kind for the cross-entropy loss.

Loss function learning is a new meta-learning paradigm that aims to automate the essential task of designing a loss function for a machine learning model. Existing techniques for loss function learning have shown promising results, often improving a model's training dynamics and final inference performance. However, a significant limitation of these techniques is that the loss functions are meta-learned in an offline fashion, where the meta-objective only considers the very first few steps of training, which is a significantly shorter time horizon than the one typically used for training deep neural networks. This causes significant bias towards loss functions that perform well at the very start of training but perform poorly at the end of training. To address this issue we propose a new loss function learning technique for adaptively updating the loss function online after each update to the base model parameters. The experimental results show that our proposed method consistently outperforms the cross-entropy loss and offline loss function learning techniques on a diverse range of neural network architectures and datasets.