Offline algorithms for Reinforcement Learning from Human Preferences (RLHF), which use only a fixed dataset of sampled responses given an input, and preference feedback among these responses, have gained increasing prominence in the literature on aligning language models. In this paper, we study how the different design choices made in methods such as DPO, IPO, SLiC and many variants influence the quality of the learned policy, from a theoretical perspective. Our treatment yields insights into the choices of loss function, the policy which is used to normalize log-likelihoods, and also the role of the data sampling policy. Notably, our results do not rely on the standard reparameterization-style arguments used to motivate some of the algorithms in this family, which allows us to give a unified treatment to a broad class of methods. We also conduct a small empirical study to verify some of the theoretical findings on a standard summarization benchmark.

The rapid proliferation of domain-specialized machine learning models presents a challenge: while individual models excel in specific domains, their performance varies significantly across diverse applications. This makes selecting the optimal model when faced with an unknown mixture of tasks, especially with limited or no data to estimate the mixture, a difficult problem. We address this challenge by formulating it as a multiple-source domain adaptation (MSA) problem. We introduce a novel, scalable algorithm that effectively routes each input to the best-suited model from a pool of available models. Our approach provides a strong performance guarantee: remarkably, for any mixture domain, the accuracy achieved by the best source model is maintained. This guarantee is established through a theoretical bound on the regret for new domains, expressed as a convex combination of the best regrets in the source domains, plus a concentration term that diminishes as the amount of source data increases. While our primary contributions are theoretical and algorithmic, we also present empirical results demonstrating the effectiveness of our approach.

We study the phenomenon of in-context learning (ICL) exhibited by large language models, where they can adapt to a new learning task, given a handful of labeled examples, without any explicit parameter optimization. Our goal is to explain how a pre-trained transformer model is able to perform ICL under reasonable assumptions on the pre-training process and the downstream tasks. We posit a mechanism whereby a transformer can achieve the following: (a) receive an i.i.d. sequence of examples which have been converted into a prompt using potentially-ambiguous delimiters, (b) correctly segment the prompt into examples and labels, (c) infer from the data a sparse linear regressor hypothesis, and finally (d) apply this hypothesis on the given test example and return a predicted label. We establish that this entire procedure is implementable using the transformer mechanism, and we give sample complexity guarantees for this learning framework. Our empirical findings validate the challenge of segmentation, and we show a correspondence between our posited mechanisms and observed attention maps for step (c).

Prod is a seminal algorithm in full-information online learning, which has been conjectured to be fundamentally sub-optimal for multi-armed bandits. By leveraging the interpretation of Prod as a first-order OMD approximation, we present the following surprising results: 1. Variants of Prod can obtain optimal regret for adversarial multi-armed bandits. 2. There exists a simple and (arguably) importance-weighting free variant with optimal rate. 3. One can even achieve best-both-worlds guarantees with logarithmic regret in the stochastic regime. The bandit algorithms in this work use simple arithmetic update rules without the need of solving optimization problems typical in prior work. Finally, the results directly improve the state of the art of incentive-compatible bandits.

In reinforcement learning applications, we often want to accurately estimate the return of several policies of interest. We study this problem, multiple-policy high-confidence policy evaluation, where the goal is to estimate the return of all given target policies up to a desired accuracy with as few samples as possible. The natural approaches to this problem, i.e., evaluating each policy separately or estimating a model of the MDP, do not take into account the similarities between target policies and scale with the number of policies to evaluate or the size of the MDP, respectively. We present an alternative approach based on reusing samples from on-policy Monte-Carlo estimators and show that it is more sample-efficient in favorable cases. Specifically, we provide guarantees in terms of a notion of overlap of the set of target policies and shed light on when such an approach is indeed beneficial compared to existing methods.

We revisit the problem of stochastic online learning with feedbackgraphs, with the goal of devising algorithms that are optimal, up toconstants, both asymptotically and in finite time. We show that,surprisingly, the notion of optimal finite-time regret is not auniquely defined property in this context and that, in general, itis decoupled from the asymptotic rate. We discuss alternativechoices and propose a notion of finite-time optimality that we argueis \emph{meaningful}. For that notion, we give an algorithm thatadmits quasi-optimal regret both in finite-time and asymptotically.

We study the problem of corralling stochastic bandit algorithms, that is combining multiple bandit algorithms designed for a stochastic environment, with the goal of devising a corralling algorithm that performs almost as well as the best base algorithm. We give two general algorithms for this setting, which we show benefit from favorable regret guarantees. We show that the regret of the corralling algorithms is no worse than that of the best algorithm containing the arm with the highest reward, and depends on the gap between the highest reward and other rewards.

We provide improved gap-dependent regret bounds for reinforcement learning in finite episodic Markov decision processes. Compared to prior work, our bounds depend on alternative definitions of gaps. These definitions are based on the insight that, in order to achieve a favorable regret, an algorithm does not need to learn how to behave optimally in states that are not reached by an optimal policy. We prove tighter upper regret bounds for optimistic algorithms and accompany them with new information-theoretic lower bounds for a large class of MDPs. Our results show that optimistic algorithms can not achieve the information-theoretic lower bounds even in deterministic MDPs unless there is a unique optimal policy.

Recent progress in model selection raises the question of the fundamental limits of these techniques. Under specific scrutiny has been model selection for general contextual bandits with nested policy classes, resulting in a COLT2020 open problem. It asks whether it is possible to obtain simultaneously the optimal single algorithm guarantees over all policies in a nested sequence of policy classes, or if otherwise this is possible for a trade-off $\alpha\in[\frac{1}{2},1)$ between complexity term and time: $\ln(|\Pi_m|)^{1-\alpha}T^\alpha$. We give a disappointing answer to this question. Even in the purely stochastic regime, the desired results are unobtainable. We present a Pareto frontier of up to logarithmic factors matching upper and lower bounds, thereby proving that an increase in the complexity term $\ln(|\Pi_m|)$ independent of $T$ is unavoidable for general policy classes.As a side result, we also resolve a COLT2016 open problem concerning second-order bounds in full-information games.

We revisit two algorithms, matrix stochastic gradient (MSG) and $\ell_2$-regularized MSG (RMSG), that are instances of stochastic gradient descent (SGD) on a convex relaxation to principal component analysis (PCA). These algorithms have been shown to outperform Oja’s algorithm, empirically, in terms of the iteration complexity, and to have runtime comparable with Oja’s. However, these findings are not supported by existing theoretical results. While the iteration complexity bound for $\ell_2$-RMSG was recently shown to match that of Oja’s algorithm, its theoretical efficiency was left as an open problem. In this work, we give improved bounds on per iteration cost of mini-batched variants of both MSG and $\ell_2$-RMSG and arrive at an algorithm with total computational complexity matching that of Oja's algorithm.

We study the adversarial multi-armed bandit problem where the learner is supplied with partial observations modeled by a \emph{feedback graph} and where shifting to a new action incurs a fixed \emph{switching cost}. We give two new algorithms for this problem in the informed setting. Our best algorithm achieves a pseudo-regret of $\tilde O(\gamma(G)^{\frac{1}{3}}T^{\frac{2}{3}})$, where $\gamma(G)$ is the domination number of the feedback graph. This significantly improves upon the previous best result for the same problem, which was based on the independence number of $G$. We also present matching lower bounds for our result that we describe in detail. Finally, we give a new algorithm with improved policy regret bounds when partial counterfactual feedback is available.

The notion of policy regret'' in online learning is supposed to capture the reactions of the adversary to the actions taken by the learner, which more traditional notions such as external regret do not take into account. We revisit this notion of policy regret, and first show that there are online learning settings in which policy regret and external regret are incompatible: any sequence of play which does well with respect to one must do poorly with respect to the other. We then focus on the game theoretic setting, when the adversary is a self-interested agent. In this setting we show that the external regret and policy regret are not in conflict, and in fact that a wide class of algorithms can ensure both as long as the adversary is also using such an algorithm. We also define a new notion of equilibrium which we call apolicy equilibrium'', and show that no-policy regret algorithms will have play which converges to such an equilibrium. Relating this back to external regret, we show that coarse correlated equilibria (which no-external regret players will converge to) are a strict subset of policy equilibria. So in game-theoretic settings every sequence of play with no external regret also has no policy regret, but the converse is not true.

We study the statistical and computational aspects of kernel principal component analysis using random Fourier features and show that under mild assumptions, $O(\sqrt{n} \log n)$ features suffices to achieve $O(1/\epsilon^2)$ sample complexity. Furthermore, we give a memory efficient streaming algorithm based on classical Oja's algorithm that achieves this rate

We propose novel first-order stochastic approximation algorithms for canonical correlation analysis (CCA). Algorithms presented are instances of inexact matrix stochastic gradient (MSG) and inexact matrix exponentiated gradient (MEG), and achieve $\epsilon$-suboptimality in the population objective in $\operatorname{poly}(\frac{1}{\epsilon})$ iterations. We also consider practical variants of the proposed algorithms and compare them with other methods for CCA both theoretically and empirically.

Partial Least Squares (PLS) is a ubiquitous statistical technique for bilinear factor analysis. It is used in many data analysis, machine learning, and information retrieval applications to model the covariance structure between a pair of data matrices. In this paper, we consider PLS for representation learning in a multiview setting where we have more than one view in data at training time. Furthermore, instead of framing PLS as a problem about a fixed given data set, we argue that PLS should be studied as a stochastic optimization problem, especially in a "big data" setting, with the goal of optimizing a population objective based on sample. This view suggests using Stochastic Approximation (SA) approaches, such as Stochastic Gradient Descent (SGD) and enables a rigorous analysis of their benefits. In this paper, we develop SA approaches to PLS and provide iteration complexity bounds for the proposed algorithms.