A common formulation of constrained reinforcement learning involves multiple rewards that must individually accumulate to given thresholds. In this class of problems, we show a simple example in which the desired optimal policy cannot be induced by any weighted linear combination of rewards. Hence, there exist constrained reinforcement learning problems for which neither regularized nor classical primal-dual methods yield optimal policies. This work addresses this shortcoming by augmenting the state with Lagrange multipliers and reinterpreting primal-dual methods as the portion of the dynamics that drives the multipliers evolution. This approach provides a systematic state augmentation procedure that is guaranteed to solve reinforcement learning problems with constraints. Thus, as we illustrate by an example, while previous methods can fail at finding optimal policies, running the dual dynamics while executing the augmented policy yields an algorithm that provably samples actions from the optimal policy.

IL is one of the mechanisms expounded by Complementary Learning Systems Theory (McClelland, McNaughton and O'Reilly, 1995; Marr, 1971) on how successful learners such as human beings mitigate effects of'catastrophic interference' while learning. Recent illustrations of IL using neural networks include Saxena, Shobe and McNaughton, 2022, who exhibited that if the new information is similar to a subset of old items, then deep neural networks can learn the new information rapidly and with the same level of accuracy by interleaving the old items in the subset. A similar insight was presented in McClelland, McNaughton and Lampinen, 2020, where it was shown that for artificial neural networks, information consistent with prior knowledge can sometimes be integrated very quickly. Another recent paper (Ban and Xie, 2021) formulated interleaved machine learning as a multi-level optimization problem, and developed an efficient differentiable algorithm to solve the interleaving learning problem with application to neural architecture search. A closely related biological concept is interleaved replay which also has been empirically validated in the literature (Gepperth and Karaoguz, 2016; Kemker and Kanan, 2018). Over the past couple of decades, ideas inspired by biological IL have been utilized in a wide array of online learning methods as well, especially to prevent catastrophic forgetting. See, for example Wang et.

We consider the problem of learning linear operators under squared loss between two infinite-dimensional Hilbert spaces in the online setting. We show that the class of linear operators with uniformly bounded $p$-Schatten norm is online learnable for any $p \in [1, \infty)$. On the other hand, we prove an impossibility result by showing that the class of uniformly bounded linear operators with respect to the operator norm is \textit{not} online learnable. Moreover, we show a separation between online uniform convergence and online learnability by identifying a class of bounded linear operators that is online learnable but uniform convergence does not hold. Finally, we prove that the impossibility result and the separation between uniform convergence and learnability also hold in the agnostic PAC setting.

Consider an online convex optimization problem where the loss functions are self-concordant barriers, smooth relative to a convex function $h$, and possibly non-Lipschitz. We analyze the regret of online mirror descent with $h$. Then, based on the result, we prove the following in a unified manner. Denote by $T$ the time horizon and $d$ the parameter dimension. 1. For online portfolio selection, the regret of $\widetilde{\text{EG}}$, a variant of exponentiated gradient due to Helmbold et al., is $\tilde{O} ( T^{2/3} d^{1/3} )$ when $T > 4 d / \log d$. This improves on the original $\tilde{O} ( T^{3/4} d^{1/2} )$ regret bound for $\widetilde{\text{EG}}$. 2. For online portfolio selection, the regret of online mirror descent with the logarithmic barrier is $\tilde{O}(\sqrt{T d})$. The regret bound is the same as that of Soft-Bayes due to Orseau et al. up to logarithmic terms. 3. For online learning quantum states with the logarithmic loss, the regret of online mirror descent with the log-determinant function is also $\tilde{O} ( \sqrt{T d} )$. Its per-iteration time is shorter than all existing algorithms we know.

Hypothesis-pruning maximizes the hypothesis updates for active learning to find those desired unlabeled data. An inherent assumption is that this learning manner can derive those updates into the optimal hypothesis. However, its convergence may not be guaranteed well if those incremental updates are negative and disordered. In this paper, we introduce a black-box teaching hypothesis $h^\mathcal{T}$ employing a tighter slack term $\left(1+\mathcal{F}^{\mathcal{T}}(\widehat{h}_t)\right)\Delta_t$ to replace the typical $2\Delta_t$ for pruning. Theoretically, we prove that, under the guidance of this teaching hypothesis, the learner can converge into a tighter generalization error and label complexity bound than those non-educated learners who do not receive any guidance from a teacher:1) the generalization error upper bound can be reduced from $R(h^*)+4\Delta_{T-1}$ to approximately $R(h^{\mathcal{T}})+2\Delta_{T-1}$, and 2) the label complexity upper bound can be decreased from $4 \theta\left(TR(h^{*})+2O(\sqrt{T})\right)$ to approximately $2\theta\left(2TR(h^{\mathcal{T}})+3 O(\sqrt{T})\right)$. To be strict with our assumption, self-improvement of teaching is firstly proposed when $h^\mathcal{T}$ loosely approximates $h^*$. Against learning, we further consider two teaching scenarios: teaching a white-box and black-box learner. Experiments verify this idea and show better generalization performance than the fundamental active learning strategies, such as IWAL, IWAL-D, etc.

Large language models (LLMs) can improve their accuracy on various tasks through iteratively refining and revising their output based on feedback. We observe that these revisions can introduce errors, in which case it is better to roll back to a previous result. Further, revisions are typically homogeneous: they use the same reasoning method that produced the initial answer, which may not correct errors. To enable exploration in this space, we present SCREWS, a modular framework for reasoning with revisions. It is comprised of three main modules: Sampling, Conditional Resampling, and Selection, each consisting of sub-modules that can be hand-selected per task. We show that SCREWS not only unifies several previous approaches under a common framework, but also reveals several novel strategies for identifying improved reasoning chains. We evaluate our framework with state-of-the-art LLMs (ChatGPT and GPT-4) on a diverse set of reasoning tasks and uncover useful new reasoning strategies for each: arithmetic word problems, multi-hop question answering, and code debugging. Heterogeneous revision strategies prove to be important, as does selection between original and revised candidates.

The widespread use of Large Language Model (LLM)-based conversational agents (CAs), especially in high-stakes domains, raises many privacy concerns. Building ethical LLM-based CAs that respect user privacy requires an in-depth understanding of the privacy risks that concern users the most. However, existing research, primarily model-centered, does not provide insight into users' perspectives. To bridge this gap, we analyzed sensitive disclosures in real-world ChatGPT conversations and conducted semi-structured interviews with 19 LLM-based CA users. We found that users are constantly faced with trade-offs between privacy, utility, and convenience when using LLM-based CAs. However, users' erroneous mental models and the dark patterns in system design limited their awareness and comprehension of the privacy risks. Additionally, the human-like interactions encouraged more sensitive disclosures, which complicated users' ability to navigate the trade-offs. We discuss practical design guidelines and the needs for paradigmatic shifts to protect the privacy of LLM-based CA users.

This work advances the field of sign language machine translation by focusing on dataset quality and simplification of the translation system. We introduce SignBank+, a clean version of the SignBank dataset, optimized for machine translation. Contrary to previous works that employ complex factorization techniques for translation, we advocate for a simplified text-to-text translation approach. Our evaluation shows that models trained on SignBank+ surpass those on the original dataset, establishing a new benchmark and providing an open resource for future research.

The problem of predicting the training time of machine learning (ML) models has become extremely relevant in the scientific community. Being able to predict a priori the training time of an ML model would enable the automatic selection of the best model both in terms of energy efficiency and in terms of performance in the context of, for instance, MLOps architectures. In this paper, we present the work we are conducting towards this direction. In particular, we present an extensive empirical study of the Full Parameter Time Complexity (FPTC) approach by Zheng et al., which is, to the best of our knowledge, the only approach formalizing the training time of ML models as a function of both dataset's and model's parameters. We study the formulations proposed for the Logistic Regression and Random Forest classifiers, and we highlight the main strengths and weaknesses of the approach. Finally, we observe how, from the conducted study, the prediction of training time is strictly related to the context (i.e., the involved dataset) and how the FPTC approach is not generalizable.

Differentiable quantum architecture search (DQAS) is a gradient-based framework to design quantum circuits automatically in the NISQ era. It was motivated by such as low fidelity of quantum hardware, low flexibility of circuit architecture, high circuit design cost, barren plateau (BP) problem, and periodicity of weights. People used it to address error mitigation, unitary decomposition, and quantum approximation optimization problems based on fixed datasets. Quantum reinforcement learning (QRL) is a part of quantum machine learning and often has various data. QRL usually uses a manually designed circuit. However, the pre-defined circuit needs more flexibility for different tasks, and the circuit design based on various datasets could become intractable in the case of a large circuit. The problem of whether DQAS can be applied to quantum deep Q-learning with various datasets is still open. The main target of this work is to discover the capability of DQAS to solve quantum deep Q-learning problems. We apply a gradient-based framework DQAS on reinforcement learning tasks and evaluate it in two different environments - cart pole and frozen lake. It contains input- and output weights, progressive search, and other new features. The experiments conclude that DQAS can design quantum circuits automatically and efficiently. The evaluation results show significant outperformance compared to the manually designed circuit. Furthermore, the performance of the automatically created circuit depends on whether the super-circuit learned well during the training process. This work is the first to show that gradient-based quantum architecture search is applicable to QRL tasks.