Haim Kaplan

Gilad Yehudai, Haim Kaplan, Guy Dar et al. · deepmind

Large language models based on the transformer architecture can solve highly complex tasks, yet their fundamental limitations on simple algorithmic problems remain poorly understood. In this work, we focus on basic counting tasks and investigate how the difficulty of these tasks scales with the transformer embedding dimension, the context length, and the vocabulary size. We reveal a sharp theoretical phase transition governed by the relationship between the embedding dimension and the vocabulary size. When the dimension is at least as large as the vocabulary, transformers can perfectly maintain token counts. However, when the vocabulary exceeds the embedding dimension, the interference between non-orthogonal token representations forces the network weights to scale polynomially. This renders the exact counting algorithm numerically unstable and practically unlearnable. We empirically validate this bottleneck by training transformers from scratch, demonstrating a strict performance drop at the theoretical threshold and catastrophic out of distribution failure when scaling the vocabulary or context length. Furthermore, we show that state-of-the-art pretrained models suffer from similar failure cases. Our work reveals a critical blind spot absent from the current literature regarding the connection among these three parameters, proving that vocabulary size fundamentally dictates the difficulty of counting tasks.

Haim Kaplan, Yishay Mansour, Shay Moran et al.

In this work we introduce an interactive variant of joint differential privacy towards handling online processes in which existing privacy definitions seem too restrictive. We study basic properties of this definition and demonstrate that it satisfies (suitable variants) of group privacy, composition, and post processing. We then study the cost of interactive joint privacy in the basic setting of online classification. We show that any (possibly non-private) learning rule can be effectively transformed to a private learning rule with only a polynomial overhead in the mistake bound. This demonstrates a stark difference with more restrictive notions of privacy such as the one studied by Golowich and Livni (2021), where only a double exponential overhead on the mistake bound is known (via an information theoretic upper bound).

Jay Tenenbaum, Haim Kaplan, Yishay Mansour et al.

We introduce the concurrent shuffle model of differential privacy. In this model we have multiple concurrent shufflers permuting messages from different, possibly overlapping, batches of users. Similarly to the standard (single) shuffle model, the privacy requirement is that the concatenation of all shuffled messages should be differentially private. We study the private continual summation problem (a.k.a. the counter problem) and show that the concurrent shuffle model allows for significantly improved error compared to a standard (single) shuffle model. Specifically, we give a summation algorithm with error $\tilde{O}(n^{1/(2k+1)})$ with $k$ concurrent shufflers on a sequence of length $n$. Furthermore, we prove that this bound is tight for any $k$, even if the algorithm can choose the sizes of the batches adaptively. For $k=\log n$ shufflers, the resulting error is polylogarithmic, much better than $\tildeΘ(n^{1/3})$ which we show is the smallest possible with a single shuffler. We use our online summation algorithm to get algorithms with improved regret bounds for the contextual linear bandit problem. In particular we get optimal $\tilde{O}(\sqrt{n})$ regret with $k= \tildeΩ(\log n)$ concurrent shufflers.

Olivier Bousquet, Haim Kaplan, Aryeh Kontorovich et al.

We construct a universally Bayes consistent learning rule that satisfies differential privacy (DP). We first handle the setting of binary classification and then extend our rule to the more general setting of density estimation (with respect to the total variation metric). The existence of a universally consistent DP learner reveals a stark difference with the distribution-free PAC model. Indeed, in the latter DP learning is extremely limited: even one-dimensional linear classifiers are not privately learnable in this stringent model. Our result thus demonstrates that by allowing the learning rate to depend on the target distribution, one can circumvent the above-mentioned impossibility result and in fact, learn \emph{arbitrary} distributions by a single DP algorithm. As an application, we prove that any VC class can be privately learned in a semi-supervised setting with a near-optimal \emph{labeled} sample complexity of $\tilde{O}(d/\varepsilon)$ labeled examples (and with an unlabeled sample complexity that can depend on the target distribution).

Yaniv Sadeh, Haim Kaplan

We study the maintenance of a $(Δ+C)$-edge-coloring ($C\ge 1$) in a fully dynamic graph $G$ with maximum degree $Δ$. We focus on minimizing \emph{recourse} which equals the number of recolored edges per edge updates. We present a new technique based on an object which we call a \emph{shift-tree}. This object tracks multiple possible recolorings of $G$ and enables us to maintain a proper coloring with small recourse in polynomial time. We shift colors over a path of edges, but unlike many other algorithms, we do not use \emph{fans} and \emph{alternating bicolored paths}. We combine the shift-tree with additional techniques to obtain an algorithm with a \emph{tight} recourse of $O\big( \frac{\log n}{\log \frac{Δ+C}{Δ-C}}\big)$ for all $C \ge 0.62Δ$ where $Δ-C = O(n^{1-δ})$. Our algorithm is the first deterministic algorithm to establish tight bounds for large palettes, and the first to do so when $Δ-C=o(Δ)$. This result settles the theoretical complexity of the recourse for large palettes. Furthermore, we believe that viewing the possible shifts as a tree can lead to similar tree-based techniques that extend to lower values of $C$, and to improved update times. A second application is to graphs with low arboricity $α$. Previous works [BCPS24, CRV24] achieve $O(ε^{-1}\log n)$ recourse per update with $C\ge (4+ε)α$, and we improve by achieving the same recourse while only requiring $C \ge (2+ε)α- 1$. This result is $Δ$-adaptive, i.e., it uses $Δ_t+C$ colors where $Δ_t$ is the current maximum degree. Trying to understand the limitations of our technique, and shift-based algorithms in general, we show a separation between the recourse achievable by algorithms that only shift colors along a path, and more general algorithms such as ones using the Nibbling Method [BGW21, BCPS24].

Anupam Gupta, Haim Kaplan, Alexander Lindermayr et al.

We simplify the proof of the optimality of the Shortest Lower-Bound First (SLF) algorithm, introduced by Gupta, Kaplan, Lindermayr, Schlöter, and Yingchareonthawornchai [FOCS'25], for minimizing the total flow time in the $\varepsilon$-clairvoyant setting.

Haim Kaplan, David Naori, Yaniv Sadeh

In the \emph{dynamic edge coloring} problem, one has to maintain a graph of maximum degree $Δ$ with at most $Δ+c$ colors, given updates to the edges of the graph. An important objective is to minimize the \emph{recourse}, which is the number of edges being recolored. We study this problem on forests, which is a natural yet nontrivial restriction of the problem. We consider the problem in both \emph{incremental} (edges are only inserted) and \emph{fully dynamic} (edges may be deleted) models. In the deterministic setting, we show that the natural greedy algorithm achieves $O(\frac{1}{c + \sqrtΔ})$ amortized recourse in the incremental model, and this is tight up to tie-breaking. In contrast, in a fully dynamic forest, greedy can be forced to have $Ω(\log_Δn)$ amortized recourse. To partially alleviate this limitation of greedy, we show an optimal non-greedy algorithm with $O(1)$ amortized recourse for \emph{rooted} fully dynamic forests and $c = Δ- 2$. In the randomized setting, we give a natural distribution-maintaining algorithm that achieves $Θ(\frac{1}Δ)$ expected amortized recourse in the incremental model and $Θ(\min \{ \fracΔ{c}, \log_Δ n \})$ expected recourse in the dynamic model. These randomized results are optimal for $c=0$.

Clara Mohri, Amir Globerson, Haim Kaplan et al.

We consider the problem of Cost-Aware Learning, where sampling different component functions of a finite-sum objective incurs different costs. The objective is to reach a target error while minimizing the total cost. First, we propose the Cost-Aware Stochastic Gradient Descent algorithm for convex functions, and derive its cost complexity to attain an error of $ε$. Furthermore, we establish a lower bound for this setting and provide a subset selection algorithm to further reduce the cost of training. We apply our theoretical insights to reinforcement learning with language models, where the computational cost of policy gradients varies with sequence length. To this end, we introduce Cost-Aware GRPO, an algorithm designed to reduce the cost of policy optimization while preserving performance. Empirical results on 1.5B and 8B LLMs demonstrate that our approach reduces the tokens used in policy optimization by up to about 30% while matching or exceeding baseline accuracy.

Marek Elias, Haim Kaplan, Yishay Mansour et al.

Recent advances in algorithmic design show how to utilize predictions obtained by machine learning models from past and present data. These approaches have demonstrated an enhancement in performance when the predictions are accurate, while also ensuring robustness by providing worst-case guarantees when predictions fail. In this paper we focus on online problems; prior research in this context was focused on a paradigm where the predictor is pre-trained on past data and then used as a black box (to get the predictions it was trained for). In contrast, in this work, we unpack the predictor and integrate the learning problem it gives rise for within the algorithmic challenge. In particular we allow the predictor to learn as it receives larger parts of the input, with the ultimate goal of designing online learning algorithms specifically tailored for the algorithmic task at hand. Adopting this perspective, we focus on a number of fundamental problems, including caching and scheduling, which have been well-studied in the black-box setting. For each of the problems we consider, we introduce new algorithms that take advantage of explicit learning algorithms which we carefully design towards optimizing the overall performance. We demonstrate the potential of our approach by deriving performance bounds which improve over those established in previous work.

Robert Busa-Fekete, Travis Dick, Claudio Gentile et al.

We investigate Learning from Label Proportions (LLP), a partial information setting where examples in a training set are grouped into bags, and only aggregate label values in each bag are available. Despite the partial observability, the goal is still to achieve small regret at the level of individual examples. We give results on the sample complexity of LLP under square loss, showing that our sample complexity is essentially optimal. From an algorithmic viewpoint, we rely on carefully designed variants of Empirical Risk Minimization, and Stochastic Gradient Descent algorithms, combined with ad hoc variance reduction techniques. On one hand, our theoretical results improve in important ways on the existing literature on LLP, specifically in the way the sample complexity depends on the bag size. On the other hand, we validate our algorithmic solutions on several datasets, demonstrating improved empirical performance (better accuracy for less samples) against recent baselines.

Clara Mohri, Haim Kaplan, Tal Schuster et al. · mit

Transformer language models generate text autoregressively, making inference latency proportional to the number of tokens generated. Speculative decoding reduces this latency without sacrificing output quality, by leveraging a small draft model to propose tokens that the larger target model verifies in parallel. In practice, however, there may exist a set of potential draft models- ranging from faster but less inaccurate, to slower yet more reliable. We introduce Hierarchical Speculative Decoding (HSD), an algorithm that stacks these draft models into a hierarchy, where each model proposes tokens, and the next larger model verifies them in a single forward pass, until finally the target model verifies tokens. We derive an expression for the expected latency of any such hierarchy and show that selecting the latency-optimal hierarchy can be done in polynomial time. Empirically, HSD gives up to 1.2x speed-up over the best single-draft baseline, demonstrating the practicality of our algorithm in reducing generation latency beyond previous techniques.

Lorne Applebaum, Travis Dick, Claudio Gentile et al.

Motivated by problems in online advertising, we address the task of Learning from Label Proportions (LLP). In this partially-supervised setting, training data consists of groups of examples, termed bags, for which we only observe the average label value. The main goal, however, remains the design of a predictor for the labels of individual examples. We introduce a novel and versatile low-variance de-biasing methodology to learn from aggregate label information, significantly advancing the state of the art in LLP. Our approach exhibits remarkable flexibility, seamlessly accommodating a broad spectrum of practically relevant loss functions across both binary and multi-class classification settings. By carefully combining our estimators with standard techniques, we substantially improve sample complexity guarantees for a large class of losses of practical relevance. We also empirically validate the efficacy of our proposed approach across a diverse array of benchmark datasets, demonstrating compelling empirical advantages over standard baselines.

Haim Kaplan, Yishay Mansour, Kobbi Nissim et al.

We introduce a new Bayesian perspective on the concept of data reconstruction, and leverage this viewpoint to propose a new security definition that, in certain settings, provably prevents reconstruction attacks. We use our paradigm to shed new light on one of the most notorious attacks in the privacy and memorization literature - fingerprinting code attacks (FPC). We argue that these attacks are really a form of membership inference attacks, rather than reconstruction attacks. Furthermore, we show that if the goal is solely to prevent reconstruction (but not membership inference), then in some cases the impossibility results derived from FPC no longer apply.

Olivier Bousquet, Amit Daniely, Haim Kaplan et al.

The amount of training-data is one of the key factors which determines the generalization capacity of learning algorithms. Intuitively, one expects the error rate to decrease as the amount of training-data increases. Perhaps surprisingly, natural attempts to formalize this intuition give rise to interesting and challenging mathematical questions. For example, in their classical book on pattern recognition, Devroye, Gyorfi, and Lugosi (1996) ask whether there exists a {monotone} Bayes-consistent algorithm. This question remained open for over 25 years, until recently Pestov (2021) resolved it for binary classification, using an intricate construction of a monotone Bayes-consistent algorithm. We derive a general result in multiclass classification, showing that every learning algorithm A can be transformed to a monotone one with similar performance. Further, the transformation is efficient and only uses a black-box oracle access to A. This demonstrates that one can provably avoid non-monotonic behaviour without compromising performance, thus answering questions asked by Devroye et al (1996), Viering, Mey, and Loog (2019), Viering and Loog (2021), and by Mhammedi (2021). Our transformation readily implies monotone learners in a variety of contexts: for example it extends Pestov's result to classification tasks with an arbitrary number of labels. This is in contrast with Pestov's work which is tailored to binary classification. In addition, we provide uniform bounds on the error of the monotone algorithm. This makes our transformation applicable in distribution-free settings. For example, in PAC learning it implies that every learnable class admits a monotone PAC learner. This resolves questions by Viering, Mey, and Loog (2019); Viering and Loog (2021); Mhammedi (2021).

Edith Cohen, Haim Kaplan, Yishay Mansour et al.

Clustering is a fundamental problem in data analysis. In differentially private clustering, the goal is to identify $k$ cluster centers without disclosing information on individual data points. Despite significant research progress, the problem had so far resisted practical solutions. In this work we aim at providing simple implementable differentially private clustering algorithms that provide utility when the data is "easy," e.g., when there exists a significant separation between the clusters. We propose a framework that allows us to apply non-private clustering algorithms to the easy instances and privately combine the results. We are able to get improved sample complexity bounds in some cases of Gaussian mixtures and $k$-means. We complement our theoretical analysis with an empirical evaluation on synthetic data.

Eliad Tsfadia, Edith Cohen, Haim Kaplan et al.

Differentially private algorithms for common metric aggregation tasks, such as clustering or averaging, often have limited practicality due to their complexity or to the large number of data points that is required for accurate results. We propose a simple and practical tool, $\mathsf{FriendlyCore}$, that takes a set of points ${\cal D}$ from an unrestricted (pseudo) metric space as input. When ${\cal D}$ has effective diameter $r$, $\mathsf{FriendlyCore}$ returns a "stable" subset ${\cal C} \subseteq {\cal D}$ that includes all points, except possibly few outliers, and is {\em certified} to have diameter $r$. $\mathsf{FriendlyCore}$ can be used to preprocess the input before privately aggregating it, potentially simplifying the aggregation or boosting its accuracy. Surprisingly, $\mathsf{FriendlyCore}$ is light-weight with no dependence on the dimension. We empirically demonstrate its advantages in boosting the accuracy of mean estimation and clustering tasks such as $k$-means and $k$-GMM, outperforming tailored methods.

Haim Kaplan, Shachar Schnapp, Uri Stemmer

In this work we study the problem of differentially private (DP) quantiles, in which given dataset $X$ and quantiles $q_1, ..., q_m \in [0,1]$, we want to output $m$ quantile estimations which are as close as possible to the true quantiles and preserve DP. We describe a simple recursive DP algorithm, which we call ApproximateQuantiles (AQ), for this task. We give a worst case upper bound on its error, and show that its error is much lower than of previous implementations on several different datasets. Furthermore, it gets this low error while running time two orders of magnitude faster that the best previous implementation.

Jay Tenenbaum, Haim Kaplan, Yishay Mansour et al.

We give an $(\varepsilon,δ)$-differentially private algorithm for the multi-armed bandit (MAB) problem in the shuffle model with a distribution-dependent regret of $O\left(\left(\sum_{a\in [k]:Δ_a>0}\frac{\log T}{Δ_a}\right)+\frac{k\sqrt{\log\frac{1}δ}\log T}{\varepsilon}\right)$, and a distribution-independent regret of $O\left(\sqrt{kT\log T}+\frac{k\sqrt{\log\frac{1}δ}\log T}{\varepsilon}\right)$, where $T$ is the number of rounds, $Δ_a$ is the suboptimality gap of the arm $a$, and $k$ is the total number of arms. Our upper bound almost matches the regret of the best known algorithms for the centralized model, and significantly outperforms the best known algorithm in the local model.

Alon Cohen, Haim Kaplan, Tomer Koren et al.

We study a novel variant of online finite-horizon Markov Decision Processes with adversarially changing loss functions and initially unknown dynamics. In each episode, the learner suffers the loss accumulated along the trajectory realized by the policy chosen for the episode, and observes aggregate bandit feedback: the trajectory is revealed along with the cumulative loss suffered, rather than the individual losses encountered along the trajectory. Our main result is a computationally efficient algorithm with $O(\sqrt{K})$ regret for this setting, where $K$ is the number of episodes. We establish this result via an efficient reduction to a novel bandit learning setting we call Distorted Linear Bandits (DLB), which is a variant of bandit linear optimization where actions chosen by the learner are adversarially distorted before they are committed. We then develop a computationally-efficient online algorithm for DLB for which we prove an $O(\sqrt{T})$ regret bound, where $T$ is the number of time steps. Our algorithm is based on online mirror descent with a self-concordant barrier regularization that employs a novel increasing learning rate schedule.

Haim Kaplan, Jay Tenenbaum

Locality Sensitive Hashing (LSH) is an effective method of indexing a set of items to support efficient nearest neighbors queries in high-dimensional spaces. The basic idea of LSH is that similar items should produce hash collisions with higher probability than dissimilar items. We study LSH for (not necessarily convex) polygons, and use it to give efficient data structures for similar shape retrieval. Arkin et al. represent polygons by their "turning function" - a function which follows the angle between the polygon's tangent and the $ x $-axis while traversing the perimeter of the polygon. They define the distance between polygons to be variations of the $ L_p $ (for $p=1,2$) distance between their turning functions. This metric is invariant under translation, rotation and scaling (and the selection of the initial point on the perimeter) and therefore models well the intuitive notion of shape resemblance. We develop and analyze LSH near neighbor data structures for several variations of the $ L_p $ distance for functions (for $p=1,2$). By applying our schemes to the turning functions of a collection of polygons we obtain efficient near neighbor LSH-based structures for polygons. To tune our structures to turning functions of polygons, we prove some new properties of these turning functions that may be of independent interest. As part of our analysis, we address the following problem which is of independent interest. Find the vertical translation of a function $ f $ that is closest in $ L_1 $ distance to a function $ g $. We prove tight bounds on the approximation guarantee obtained by the translation which is equal to the difference between the averages of $ g $ and $ f $.

Haim Kaplan, Yishay Mansour, Uri Stemmer

We revisit one of the most basic and widely applicable techniques in the literature of differential privacy - the sparse vector technique [Dwork et al., STOC 2009]. This simple algorithm privately tests whether the value of a given query on a database is close to what we expect it to be. It allows to ask an unbounded number of queries as long as the answer is close to what we expect, and halts following the first query for which this is not the case. We suggest an alternative, equally simple, algorithm that can continue testing queries as long as any single individual does not contribute to the answer of too many queries whose answer deviates substantially form what we expect. Our analysis is subtle and some of its ingredients may be more widely applicable. In some cases our new algorithm allows to privately extract much more information from the database than the original. We demonstrate this by applying our algorithm to the shifting heavy-hitters problem: On every time step, each of $n$ users gets a new input, and the task is to privately identify all the current heavy-hitters. That is, on time step $i$, the goal is to identify all data elements $x$ such that many of the users have $x$ as their current input. We present an algorithm for this problem with improved error guarantees over what can be obtained using existing techniques. Specifically, the error of our algorithm depends on the maximal number of times that a single user holds a heavy-hitter as input, rather than the total number of times in which a heavy-hitter exists.

Haim Kaplan, Yishay Mansour, Uri Stemmer et al.

We present a differentially private learner for halfspaces over a finite grid $G$ in $\mathbb{R}^d$ with sample complexity $\approx d^{2.5}\cdot 2^{\log^*|G|}$, which improves the state-of-the-art result of [Beimel et al., COLT 2019] by a $d^2$ factor. The building block for our learner is a new differentially private algorithm for approximately solving the linear feasibility problem: Given a feasible collection of $m$ linear constraints of the form $Ax\geq b$, the task is to privately identify a solution $x$ that satisfies most of the constraints. Our algorithm is iterative, where each iteration determines the next coordinate of the constructed solution $x$.

Haim Kaplan, Jay Tenenbaum

Locality Sensitive Hashing (LSH) is an effective method to index a set of points such that we can efficiently find the nearest neighbors of a query point. We extend this method to our novel Set-query LSH (SLSH), such that it can find the nearest neighbors of a set of points, given as a query. Let $ s(x,y) $ be the similarity between two points $ x $ and $ y $. We define a similarity between a set $ Q$ and a point $ x $ by aggregating the similarities $ s(p,x) $ for all $ p\in Q $. For example, we can take $ s(p,x) $ to be the angular similarity between $ p $ and $ x $ (i.e., $1-{\angle (x,p)}/π$), and aggregate by arithmetic or geometric averaging, or taking the lowest similarity. We develop locality sensitive hash families and data structures for a large set of such arithmetic and geometric averaging similarities, and analyze their collision probabilities. We also establish an analogous framework and hash families for distance functions. Specifically, we give a structure for the euclidean distance aggregated by either averaging or taking the maximum. We leverage SLSH to solve a geometric extension of the approximate near neighbors problem. In this version, we consider a metric for which the unit ball is an ellipsoid and its orientation is specified with the query. An important application that motivates our work is group recommendation systems. Such a system embeds movies and users in the same feature space, and the task of recommending a movie for a group to watch together, translates to a set-query $ Q $ using an appropriate similarity.

Avinatan Hassidim, Haim Kaplan, Yishay Mansour et al.

A streaming algorithm is said to be adversarially robust if its accuracy guarantees are maintained even when the data stream is chosen maliciously, by an adaptive adversary. We establish a connection between adversarial robustness of streaming algorithms and the notion of differential privacy. This connection allows us to design new adversarially robust streaming algorithms that outperform the current state-of-the-art constructions for many interesting regimes of parameters.

Haim Kaplan, Micha Sharir, Uri Stemmer

We study the question of how to compute a point in the convex hull of an input set $S$ of $n$ points in ${\mathbb R}^d$ in a differentially private manner. This question, which is trivial non-privately, turns out to be quite deep when imposing differential privacy. In particular, it is known that the input points must reside on a fixed finite subset $G\subseteq{\mathbb R}^d$, and furthermore, the size of $S$ must grow with the size of $G$. Previous works focused on understanding how $n$ needs to grow with $|G|$, and showed that $n=O\left(d^{2.5}\cdot8^{\log^*|G|}\right)$ suffices (so $n$ does not have to grow significantly with $|G|$). However, the available constructions exhibit running time at least $|G|^{d^2}$, where typically $|G|=X^d$ for some (large) discretization parameter $X$, so the running time is in fact $Ω(X^{d^3})$. In this paper we give a differentially private algorithm that runs in $O(n^d)$ time, assuming that $n=Ω(d^4\log X)$. To get this result we study and exploit some structural properties of the Tukey levels (the regions $D_{\ge k}$ consisting of points whose Tukey depth is at least $k$, for $k=0,1,...$). In particular, we derive lower bounds on their volumes for point sets $S$ in general position, and develop a rather subtle mechanism for handling point sets $S$ in degenerate position (where the deep Tukey regions have zero volume). A naive approach to the construction of the Tukey regions requires $n^{O(d^2)}$ time. To reduce the cost to $O(n^d)$, we use an approximation scheme for estimating the volumes of the Tukey regions (within their affine spans in case of degeneracy), and for sampling a point from such a region, a scheme that is based on the volume estimation framework of Lovász and Vempala (FOCS 2003) and of Cousins and Vempala (STOC 2015). Making this framework differentially private raises a set of technical challenges that we address.

Alon Cohen, Haim Kaplan, Yishay Mansour et al.

Stochastic shortest path (SSP) is a well-known problem in planning and control, in which an agent has to reach a goal state in minimum total expected cost. In the learning formulation of the problem, the agent is unaware of the environment dynamics (i.e., the transition function) and has to repeatedly play for a given number of episodes while reasoning about the problem's optimal solution. Unlike other well-studied models in reinforcement learning (RL), the length of an episode is not predetermined (or bounded) and is influenced by the agent's actions. Recently, Tarbouriech et al. (2019) studied this problem in the context of regret minimization and provided an algorithm whose regret bound is inversely proportional to the square root of the minimum instantaneous cost. In this work we remove this dependence on the minimum cost---we give an algorithm that guarantees a regret bound of $\widetilde{O}(B_\star |S| \sqrt{|A| K})$, where $B_\star$ is an upper bound on the expected cost of the optimal policy, $S$ is the set of states, $A$ is the set of actions and $K$ is the number of episodes. We additionally show that any learning algorithm must have at least $Ω(B_\star \sqrt{|S| |A| K})$ regret in the worst case.

Haim Kaplan, Katrina Ligett, Yishay Mansour et al.

We study the sample complexity of learning threshold functions under the constraint of differential privacy. It is assumed that each labeled example in the training data is the information of one individual and we would like to come up with a generalizing hypothesis $h$ while guaranteeing differential privacy for the individuals. Intuitively, this means that any single labeled example in the training data should not have a significant effect on the choice of the hypothesis. This problem has received much attention recently; unlike the non-private case, where the sample complexity is independent of the domain size and just depends on the desired accuracy and confidence, for private learning the sample complexity must depend on the domain size $X$ (even for approximate differential privacy). Alon et al. (STOC 2019) showed a lower bound of $Ω(\log^*|X|)$ on the sample complexity and Bun et al. (FOCS 2015) presented an approximate-private learner with sample complexity $\tilde{O}\left(2^{\log^*|X|}\right)$. In this work we reduce this gap significantly, almost settling the sample complexity. We first present a new upper bound (algorithm) of $\tilde{O}\left(\left(\log^*|X|\right)^2\right)$ on the sample complexity and then present an improved version with sample complexity $\tilde{O}\left(\left(\log^*|X|\right)^{1.5}\right)$. Our algorithm is constructed for the related interior point problem, where the goal is to find a point between the largest and smallest input elements. It is based on selecting an input-dependent hash function and using it to embed the database into a domain whose size is reduced logarithmically; this results in a new database, an interior point of which can be used to generate an interior point in the original database in a differentially private manner.

Tom Zahavy, Alon Cohen, Haim Kaplan et al.

We consider the applications of the Frank-Wolfe (FW) algorithm for Apprenticeship Learning (AL). In this setting, we are given a Markov Decision Process (MDP) without an explicit reward function. Instead, we observe an expert that acts according to some policy, and the goal is to find a policy whose feature expectations are closest to those of the expert policy. We formulate this problem as finding the projection of the feature expectations of the expert on the feature expectations polytope -- the convex hull of the feature expectations of all the deterministic policies in the MDP. We show that this formulation is equivalent to the AL objective and that solving this problem using the FW algorithm is equivalent well-known Projection method of Abbeel and Ng (2004). This insight allows us to analyze AL with tools from convex optimization literature and derive tighter convergence bounds on AL. Specifically, we show that a variation of the FW method that is based on taking "away steps" achieves a linear rate of convergence when applied to AL and that a stochastic version of the FW algorithm can be used to avoid precise estimation of feature expectations. We also experimentally show that this version outperforms the FW baseline. To the best of our knowledge, this is the first work that shows linear convergence rates for AL.

Gal Sadeh, Edith Cohen, Haim Kaplan

Influence maximization (IM) is the problem of finding for a given $s\geq 1$ a set $S$ of $|S|=s$ nodes in a network with maximum influence. With stochastic diffusion models, the influence of a set $S$ of seed nodes is defined as the expectation of its reachability over simulations, where each simulation specifies a deterministic reachability function. Two well-studied special cases are the Independent Cascade (IC) and the Linear Threshold (LT) models of Kempe, Kleinberg, and Tardos. The influence function in stochastic diffusion is unbiasedly estimated by averaging reachability values over i.i.d. simulations. We study the IM sample complexity: the number of simulations needed to determine a $(1-ε)$-approximate maximizer with confidence $1-δ$. Our main result is a surprising upper bound of $O( s τε^{-2} \ln \frac{n}δ)$ for a broad class of models that includes IC and LT models and their mixtures, where $n$ is the number of nodes and $τ$ is the number of diffusion steps. Generally $τ\ll n$, so this significantly improves over the generic upper bound of $O(s n ε^{-2} \ln \frac{n}δ)$. Our sample complexity bounds are derived from novel upper bounds on the variance of the reachability that allow for small relative error for influential sets and additive error when influence is small. Moreover, we provide a data-adaptive method that can detect and utilize fewer simulations on models where it suffices. Finally, we provide an efficient greedy design that computes an $(1-1/e-ε)$-approximate maximizer from simulations and applies to any submodular stochastic diffusion model that satisfies the variance bounds.

Yuval Lewi, Haim Kaplan, Yishay Mansour

We study the Thompson sampling algorithm in an adversarial setting, specifically, for adversarial bit prediction. We characterize the bit sequences with the smallest and largest expected regret. Among sequences of length $T$ with $k < \frac{T}{2}$ zeros, the sequences of largest regret consist of alternating zeros and ones followed by the remaining ones, and the sequence of smallest regret consists of ones followed by zeros. We also bound the regret of those sequences, the worse case sequences have regret $O(\sqrt{T})$ and the best case sequence have regret $O(1)$. We extend our results to a model where false positive and false negative errors have different weights. We characterize the sequences with largest expected regret in this generalized setting, and derive their regret bounds. We also show that there are sequences with $O(1)$ regret.

Tom Zahavy, Alon Cohen, Haim Kaplan et al.

We derive and analyze learning algorithms for apprenticeship learning, policy evaluation, and policy gradient for average reward criteria. Existing algorithms explicitly require an upper bound on the mixing time. In contrast, we build on ideas from Markov chain theory and derive sampling algorithms that do not require such an upper bound. For these algorithms, we provide theoretical bounds on their sample-complexity and running time.

Tom Zahavy, Avinatan Hasidim, Haim Kaplan et al.

We consider a settings of hierarchical reinforcement learning, in which the reward is a sum of components. For each component we are given a policy that maximizes it and our goal is to assemble a policy from the individual policies that maximizes the sum of the components. We provide theoretical guarantees for assembling such policies in deterministic MDPs with collectible rewards. Our approach builds on formulating this problem as a traveling salesman problem with discounted reward. We focus on local solutions, i.e., policies that only use information from the current state; thus, they are easy to implement and do not require substantial computational resources. We propose three local stochastic policies and prove that they guarantee better performance than any deterministic local policy in the worst case; experimental results suggest that they also perform better on average.

Haim Kaplan, Yishay Mansour, Yossi Matias et al.

We present differentially private efficient algorithms for learning union of polygons in the plane (which are not necessarily convex). Our algorithms achieve $(α,β)$-PAC learning and $(ε,δ)$-differential privacy using a sample of size $\tilde{O}\left(\frac{1}{αε}k\log d\right)$, where the domain is $[d]\times[d]$ and $k$ is the number of edges in the union of polygons.

Alon Cohen, Avinatan Hassidim, Haim Kaplan et al.

Imagine a large firm with multiple departments that plans a large recruitment. Candidates arrive one-by-one, and for each candidate the firm decides, based on her data (CV, skills, experience, etc), whether to summon her for an interview. The firm wants to recruit the best candidates while minimizing the number of interviews. We model such scenarios as an assignment problem between items (candidates) and categories (departments): the items arrive one-by-one in an online manner, and upon processing each item the algorithm decides, based on its value and the categories it can be matched with, whether to retain or discard it (this decision is irrevocable). The goal is to retain as few items as possible while guaranteeing that the set of retained items contains an optimal matching. We consider two variants of this problem: (i) in the first variant it is assumed that the $n$ items are drawn independently from an unknown distribution $D$. (ii) In the second variant it is assumed that before the process starts, the algorithm has an access to a training set of $n$ items drawn independently from the same unknown distribution (e.g.\ data of candidates from previous recruitment seasons). We give tight bounds on the minimum possible number of retained items in each of these variants. These results demonstrate that one can retain exponentially less items in the second variant (with the training set).

Tom Zahavy, Avinatan Hasidim, Haim Kaplan et al.

In this work, we provide theoretical guarantees for reward decomposition in deterministic MDPs. Reward decomposition is a special case of Hierarchical Reinforcement Learning, that allows one to learn many policies in parallel and combine them into a composite solution. Our approach builds on mapping this problem into a Reward Discounted Traveling Salesman Problem, and then deriving approximate solutions for it. In particular, we focus on approximate solutions that are local, i.e., solutions that only observe information about the current state. Local policies are easy to implement and do not require substantial computational resources as they do not perform planning. While local deterministic policies, like Nearest Neighbor, are being used in practice for hierarchical reinforcement learning, we propose three stochastic policies that guarantee better performance than any deterministic policy.

Edith Cohen, Shiri Chechik, Haim Kaplan

Clustering of data points is a fundamental tool in data analysis. We consider points $X$ in a relaxed metric space, where the triangle inequality holds within a constant factor. The {\em cost} of clustering $X$ by $Q$ is $V(Q)=\sum_{x\in X} d_{xQ}$. Two basic tasks, parametrized by $k \geq 1$, are {\em cost estimation}, which returns (approximate) $V(Q)$ for queries $Q$ such that $|Q|=k$ and {\em clustering}, which returns an (approximate) minimizer of $V(Q)$ of size $|Q|=k$. With very large data sets $X$, we seek efficient constructions of small samples that act as surrogates to the full data for performing these tasks. Existing constructions that provide quality guarantees are either worst-case, and unable to benefit from structure of real data sets, or make explicit strong assumptions on the structure. We show here how to avoid both these pitfalls using adaptive designs. At the core of our design is the {\em one2all} construction of multi-objective probability-proportional-to-size (pps) samples: Given a set $M$ of centroids and $α\geq 1$, one2all efficiently assigns probabilities to points so that the clustering cost of {\em each} $Q$ with cost $V(Q) \geq V(M)/α$ can be estimated well from a sample of size $O(α|M|ε^{-2})$. For cost queries, we can obtain worst-case sample size $O(kε^{-2})$ by applying one2all to a bicriteria approximation $M$, but we adaptively balance $|M|$ and $α$ to further reduce sample size. For clustering, we design an adaptive wrapper that applies a base clustering algorithm to a sample $S$. Our wrapper uses the smallest sample that provides statistical guarantees that the quality of the clustering on the sample carries over to the full data set. We demonstrate experimentally the huge gains of using our adaptive instead of worst-case methods.

Shiri Chechik, Edith Cohen, Haim Kaplan

The average distance from a node to all other nodes in a graph, or from a query point in a metric space to a set of points, is a fundamental quantity in data analysis. The inverse of the average distance, known as the (classic) closeness centrality of a node, is a popular importance measure in the study of social networks. We develop novel structural insights on the sparsifiability of the distance relation via weighted sampling. Based on that, we present highly practical algorithms with strong statistical guarantees for fundamental problems. We show that the average distance (and hence the centrality) for all nodes in a graph can be estimated using $O(ε^{-2})$ single-source distance computations. For a set $V$ of $n$ points in a metric space, we show that after preprocessing which uses $O(n)$ distance computations we can compute a weighted sample $S\subset V$ of size $O(ε^{-2})$ such that the average distance from any query point $v$ to $V$ can be estimated from the distances from $v$ to $S$. Finally, we show that for a set of points $V$ in a metric space, we can estimate the average pairwise distance using $O(n+ε^{-2})$ distance computations. The estimate is based on a weighted sample of $O(ε^{-2})$ pairs of points, which is computed using $O(n)$ distance computations. Our estimates are unbiased with normalized mean square error (NRMSE) of at most $ε$. Increasing the sample size by a $O(\log n)$ factor ensures that the probability that the relative error exceeds $ε$ is polynomially small.