EnCORE

Abstract:

We give an efficient reduction through which any machine learning algorithm can be converted into an interactive protocol that can interact with a human decision maker to improve their predictions. In the interactive protocol, the machine learning model first produces a prediction, which implies a recommended action for the human decision maker. The human can update their beliefs as a function of the recommendation they received, but can also incorporate their own knowledge and observations, and so may disagree with the provided recommendation. The human then responds to the model’s recommendation by conveying either agreement with the model’s recommendation, or directional disagreement. The model then updates its state and provides a new recommendation, and the human may in turn again update their beliefs given the new information they have learned from the model. The process continues until the model and the human reach agreement. We show that any predictive model can be efficiently turned into an agreement protocol (without reducing its accuracy) so that the number of rounds until the protocol reaches agreement is small, under forgiving, computationally tractable assumptions on the human’s decision making process. Additionally, the final decisions are guaranteed to be more accurate than either those of the model or of the human on their own. In fact, the assumptions we make on the human would be satisfied by a Bayesian decision maker, but are a substantial relaxation that makes them algorithmically tractable and do not require perfect Bayesian rationality. In this way we generalize Aumann’s agreement theorem from requiring Bayesian rationality to requiring only tractable calibration conditions, and generalize known quantitative versions of the theorem to hold for multi-dimensional expectations and more general forms of communication.

Abstract:

Machine learning models are increasingly used in societal applications, yet legal and privacy concerns demand that they very often be kept confidential. Consequently, there is a growing distrust about the fairness properties of these models in the minds of consumers, who are often at the receiving end of model predictions. To this end, we propose Fairproof – a system that uses Zero-Knowledge Proofs (a cryptographic primitive) to publicly verify the fairness of a model, while maintaining confidentiality. We also propose a fairness certification algorithm for fully-connected neural networks which is befitting to ZKPs and is used in this system. We implement Fairproof in Gnark and demonstrate empirically that our system is practically feasible.. 

Abstract: How do we build ML systems that put the interests of users and society front and center? Using collaborations with Doctors Without Borders and the Cancer Registry of Norway as case studies, we will show how collaborative learning (CL) can be used to build such user-centric ML systems. CL techniques (including decentralized and federated learning) allow us to train ML models without the users sacrificing their privacy or ownership and control over their data.  Yet for these systems to truly succeed, two fundamental challenges must be confronted. These systems need to be efficient and scale to massive networks, and manage and resolve the conflicting goals of the participants. We will discuss how tools from optimization, statistics, and economics can be leveraged to address these challenges.

Bio: Sai Praneeth Karimireddy is a postdoc at UC Berkeley with Mike I. Jordan and obtained his PhD from EPFL with Martin Jaggi. His research builds large-scale collaborative learning systems and has seen widespread real-world adoption both by public health organizations (e.g., Doctors Without Borders, the Red Cross, the Cancer Registry of Norway) and by industries such as Meta, Google, OpenAI, and Owkin. His research has been recognized by the EPFL Patrick Denantes Memorial Prize for the best thesis in computer science, the Chorafas Foundation Award for exceptional applied research, an EPFL thesis distinction award, an SNSF fellowship, and multiple best paper awards.

Abstract: Distributed high dimensional mean estimation is a common aggregation routine used often in distributed optimization methods (e.g. federated learning). In light of these applications, recently there has been an interest in distributed mean estimation in a communication-constrained setting, where vectors have to be compressed before sharing. However,  in these applications the vectors whose mean are to be estimated are often correlated with each other. To exploit these correlations, recently Suresh et al., 2022, Jhunjhunwala et al., 2021, Jiang et al, 2023, proposed multiple  correlation aware compression schemes. However, any theoretical analysis of graceful degradation of these correlation-aware compression schemes with increasing heterogeneity (de-correlation) is absent in  the literature. Moreover, the correlations have to be known for these schemes to work. In this paper, we propose collaborative compression schemes that agnostically exploit the correlation among vectors in a distributed setting. We propose modifications of 4 different compression schemes to make them suitable for distributed mean estimation. Our schemes are all based on and/or variants of popular sign-based compression (signSGD and 1 bit compressed sensing). We do a rigorous theoretical analysis of our proposed schemes to show how the estimation error varies with the degree of correlation among vectors. In the process, we come up with appropriate correlation measures for these applications as well. Further, we compare the performance of our schemes to existing benchmarks for distributed mean estimation and for downstream distributed learning tasks.

Abstract: In federated learning, a large number of local steps is usually avoided due to data heterogeneity. However, in this work we leverage the implicit bias of gradient descent to show that a federated approach involving a large number of local steps can closely approximate the centralized solution obtained by training on aggregated data. We investigate  this phenomenon through experiments conducted on synthetic data using both linear and logistic regression models. Furthermore, we offer a theoretical explanation for the linear regression task.

Abstract: In many dimensionality reduction tasks, we wish to identify the constituent components that explain our observations. For manifold learning, this can be formalized as factoring a Riemannian product manifold. Recovering this factorization, however, may suffer from certain difficulties in practice, especially when data is sparse or noisy, or when one factor is distorted by the other. To address these limitations, we propose identifying non-redundant coordinates on the product manifold before applying product manifold learning to identify which coordinates correspond to different factor manifolds.

Abstract: We present a framework for aligning the local views of a possibly closed/non-orientable data manifold to produce an embedding in its intrinsic dimension through tearing. Through a spectral coloring scheme, we render the embeddings of the points across the tear with matching colors, enabling a visual recovery of the topology of the data manifold. The embedding is further equipped with a tear-aware metric that enables computation of shortest paths while accounting for the tear. To measure the quality of an embedding, we propose two Lipschitz-type notions of global distortion—a stronger and a weaker one—along with their pointwise counterparts for a finer assessment of the embedding. Subsequently, we bound them using the distortion of the local views and the alignment error between them. We show that our theoretical result on strong distortion leads to a new perspective on the need for a repulsion term in manifold learning objectives. As a result, we enhance our alignment approach by incorporating repulsion. Finally, we compare various strategies for the tear and repulsion enabled alignment, with regard to their speed of convergence and the quality of the embeddings produced.

This is joint work with my advisors Gal Mishne and Alex Cloninger at UCSD.

Abstract: In this talk, I will consider an online-learning problem, called Online Clustering with Moving Costs, at which a learner maintains a set of facilities over rounds so as to minimize the connection cost of an adversarially selected sequence of clients. The learner is informed on the positions of the clients at each round only after its facility-selection and can use this information to update its decision in the next round. However, updating the facility positions comes with an additional moving cost based on the moving distance of the facilities. I will be presenting the first (polynomial-time) approximate-regret algorithm for this setting through a combination of different algorithmic techniques such as HST embeddings, the FTRL framework with a dilated entropic regulariser as well as a novel rounding scheme.

Abstract: Despite the success of Transformers on language understanding, code generation, and logical reasoning, they still fail to (length) generalize on basic arithmetic tasks such as addition and multiplication. A major reason behind this failure is the vast difference in structure of numbers versus text; for example, numbers are associated with a specific significance order that plays a role in calculating the answer. In contrast, for text, such symmetries are quite unnatural. In this work, we propose to encode these semantics explicitly into the model via appropriate data formatting and custom positional encodings. To further elucidate the importance of explicitly encoding structure, in a simplified linear setting, we prove that standard positional encodings even when trained with augmentations to implicitly induce structure fail at such generalization, whereas enforcing structure via positional encodings succeeds.

Bio: Mahdi Sabbaghi is a second year PhD student at UPenn, department of Electrical and System Engineering, supervised by Professors Hamed Hassani and George Pappas. Previously he obtained a B. Sc. degree in Electrical Engineering as well as a B. Sc. degree in Physics from the Sharif University of Technology, in Tehran.

Abstract: We consider the problem of private estimation of U statistics. U statistics are widely used estimators that naturally arise in a broad class of problems, from nonparametric signed rank tests to subgraph counts in random networks. Despite the recent outpouring of interest in private mean estimation, private algorithms for more general U statistics have received little attention.  We propose a framework where, for a broad class of U statistics, one can use existing tools in private mean estimation to obtain confidence intervals where the private error does not overwhelm the irreducible error resulting from the variance of the U statistics. However, in specific cases that arise when the U statistics degenerate or have vanishing moments, the private error may be of a larger order than the non-private error. To remedy this, we propose a new thresholding-based approach that uses Hajek projections to re-weight different subsets. As we show,  this leads to more accurate inference in certain settings.

Abstract: We derive the first finite-time logarithmic Bayes regret upper bounds for Bayesian bandits, for BayesUCB and Thompson Sampling. In Gaussian and Bernoulli multi-armed bandits, we obtain $O(c_\Delta \log n)$ and $O(c_h \log^2 n)$ upper bounds for an upper confidence bound algorithm, where $c_h$ and $c_\Delta$ are constants depending on the prior distribution and the gaps of bandit instances sampled from it, respectively. The latter bound asymptotically matches the lower bound of Lai (1987). Our proofs are a major technical departure from prior works, while being simple and general. The key idea in our proofs is to conduct a frequentist per-instance analysis with Bayesian confidence intervals, and then integrate it over the prior.

Our results provide insights on the value of prior in the Bayesian setting, both in the objective and as a side information given to the learner. They significantly improve upon existing $\tilde{O}(\sqrt{n})$ bounds, which have become standard in the literature despite the logarithmic lower bound of Lai (1987).

Abstract: We study the problem of characterizing optimal learning algorithms for playing repeated games against an adversary with unknown payoffs. In this problem, the first player (called the learner) commits to a learning algorithm against a second player (called the optimizer), and the optimizer best-responds by choosing the optimal dynamic strategy for their (unknown but well-defined) payoff. Classic learning algorithms (such as no-regret algorithms) provide some counterfactual guarantees for the learner, but might perform much more poorly than other learning algorithms against particular optimizer payoffs.

In this paper, we introduce the notion of asymptotically Pareto-optimal learning algorithms. Intuitively, if a learning algorithm is Pareto-optimal, then there is no other algorithm which performs asymptotically at least as well against all optimizers and performs strictly better (by at least $\Omega(T)$) against some optimizer. We show that well-known no-regret algorithms such as Multiplicative Weights and Follow The Regularized Leader are Pareto-dominated. However, while no-regret is not enough to ensure Pareto-optimality, we show that a strictly stronger property, no-swap-regret, is a sufficient condition for Pareto-optimality.

Proving these results requires us to address various technical challenges specific to repeated play, including the fact that there is no simple characterization of how optimizers who are rational in the long-term best-respond against a learning algorithm over multiple rounds of play. To address this, we introduce the idea of the asymptotic menu of a learning algorithm: the convex closure of all correlated distributions over strategy profiles that are asymptotically implementable by an adversary. Interestingly, we show that all no-swap-regret algorithms share the same asymptotic menu, implying that all no-swap-regret algorithms are “strategically equivalent”.

This talk is based on work with Jon Schneider.

Abstract: We consider crowd-based metric learning from preference comparisons, where given two items, a user prefers the item that is closer to their latent ideal item. Here, “closeness” is measured with respect to a shared but unknown Mahalanobis distance. Can we recover this distance when we can only obtain very few responses per user?

In this very low-budget regime, we show that generally, nothing at all about the metric is revealed, even with infinitely many users. But when the items have subspace cluster structure, we present a divide-and-conquer approach for metric recovery, and provide theoretical recovery guarantees and empirical validation.

This is joint work with Zhi Wang (UCSD) and Ramya Korlakai Vinayak (UW–Madison).

Abstract: We investigate the concept of effective resistance in connection graphs, expanding its traditional application from undirected graphs. We propose a robust definition of effective resistance in connection graphs by focusing on the duality of Dirichlet-type and Poisson-type problems on connection graphs. Additionally, we delve into random walks, taking into account both node transitions and vector rotations. This approach introduces novel concepts of effective conductance and resistance matrices for connection graphs, capturing mean rotation matrices corresponding to random walk transitions. Thereby, it provides new theoretical insights for network analysis and optimization.

This is based on a joint work with Alexander Cloninger, Gal Mishne, Andreas Oslandsbotn, Sawyer Jack Robertson and Yusu Wang.

Abstract: We study the approximability and inapproximability of Strict-CSPs. An instance of the Strict-CSPs consists of a set of constraints over a set of variables and a cost function over the assignments. The goal is to find an assignment to the variables of minimum cost which satisfies all the constraints. Some prominent problems that this framework captures are (Hypergraph) Vertex Cover, Min Sum k-Coloring, Multiway Cut, Min Ones, and others.

We focus on a systematic study of Strict-CSPs of the form Strict-CSPs(H), that is, Strict-CSPs where the type of constraints is limited to predicates from a set H. Our first result is a dichotomy for approximation of Strict-CSPs(H), where H is a binary predicate, i.e., a digraph. We prove that if digraph H has bounded width, then Strict-CSPs(H) is approximable within a constant factor (depending on H); otherwise, there is no approximation for Strict-CSPs(H) unless P=NP.

Second, we study the inapproximability of Strict-CSP and present the first general hardness of approximation for Strict-CSP. More precisely, we prove a dichotomy theorem that states every instance of Strict-CSP(H) (H being a digraph) is either polynomial-time solvable or APX-complete. Moreover, we show the existence of a universal constant 0<\delta<1 such that it is NP-hard to approximate Strict-CSP(H) within a factor of (2-\delta) for all digraphs H where Strict-CSP(H) is NP-complete.

Abstract: A recent line of research has established a novel desideratum for designing approximatelyrevenue-optimal multi-item mechanisms, namely the buy-many constraint. Under this constraint, prices for different allocations made by the mechanism must be subadditive implying that the price of a bundle cannot exceed the sum of prices of individual items it contains. This natural constraint has enabled several positive results in multi-item mechanism design bypassing well-established impossibility results. Our work addresses a main open question from this literature involving the design of buymany mechanisms for multiple buyers. Our main result is that a simple sequential item pricing mechanism with buyer-specific prices can achieve an O(log m) approximation to the revenue of any buy-many mechanism when all buyers have unit-demand preferences over m items. This is the best possible as it directly matches the previous results for the single-buyer setting where no simple mechanism can obtain a better approximation. Our result applies in full generality: even though there are many alternative ways buy-many mechanisms can be defined for multibuyer settings, our result captures all of them at the same time. We achieve this by directly competing with a more permissive upper-bound on the buy-many revenue, obtained via an ex-ante relaxation.

Abstract: Since its inception in 1982, Oja’s algorithm has become an established method for streaming principle component analysis (PCA). We study the problem of streaming PCA, where the data-points are sampled from an irreducible, aperiodic, and reversible Markov chain. Our goal is to estimate the top eigenvector of the unknown covariance matrix of the stationary distribution. This setting has implications in scenarios where data can solely be sampled from a Markov Chain Monte Carlo (MCMC) type algorithm, and the objective is to perform inference on parameters of the stationary distribution. Most convergence guarantees for Oja’s algorithm in the literature assume that the data-points are sampled IID. For data streams with Markovian dependence, one typically downsamples the data to get a “nearly” independent data stream. In this paper, we obtain the first sharp rate for Oja’s algorithm on the entire data, where we remove the logarithmic dependence on the sample size, resulting from throwing data away in downsampling strategies.

Abstract: Monotonicity testing of Boolean functions over the n-width, d-dimensional hypergrid is a classic problem in property testing, where the goal is to design a randomized algorithm which can distinguish monotone functions from those which are far from any monotone function while making as few queries as possible. The special case of n = 2 corresponds to the hypercube domain. Here a long line of works exploiting a very interesting connection with isoperimetric inequalities for Boolean functions culminated in a non-adaptive tester making ~O(d^{1/2}) queries in a celebrated paper by Khot, Minzer, and Safra (SICOMP 2018). This is known to be optimal for non-adaptive testers. However, the general case of hypergrids for n > 2 remained open. Very recently, two papers (Black-Chakrabarty-Seshadhri STOC 2023 and Braverman-Khot-Kindler-Minzer ITCS 2023) independently obtained ~O(poly(n) d^{1/2}) query testers for hypergrids. These results are essentially optimal for n < polylog(d), but are far from optimal for n >> polylog(d).

This talk covers our most recent result (appearing at FOCS 2023) which obtains a non-adaptive d^{1/2+o(1)} query tester for all n, resolving the non-adaptive monotonicity testing problem for hypergrids, up to a factor of d^{o(1)}. Our proof relies on many new techniques as well as two key theorems which we proved in earlier works from SODA 2020 and STOC 2023.

Abstract: Despite efforts to align large language models (LLMs) with human values, widely-used LLMs such as GPT, Llama, Claude, and PaLM are susceptible to jailbreaking attacks, wherein an adversary fools a targeted LLM into generating objectionable content.  To address this vulnerability, we propose SmoothLLM, the first algorithm designed to mitigate jailbreaking attacks on LLMs.  Based on our finding that adversarially-generated prompts are brittle to character-level changes, our defense first randomly perturbs multiple copies of a given input prompt, and then aggregates the corresponding predictions to detect adversarial inputs.  SmoothLLM reduces the attack success rate on numerous popular LLMs to below one percentage point, avoids unnecessary conservatism, and admits provable guarantees on attack mitigation.  Moreover, our defense uses exponentially fewer queries than existing attacks and is compatible with any LLM.

Abstract: We study the design of embeddings into Euclidean space with outliers. Given a metric space $(X,d)$ and an integer $k$, the goal is to embed all but $k$ points in $X$ (called the “”outliers””) into $\ell_2$ with the smallest possible distortion $c$. Finding the optimal distortion $c$ for a given outlier set size $k$, or alternately the smallest $k$ for a given target distortion $c$ are both NP-hard problems. In fact, it is UGC-hard to approximate $k$ to within a factor smaller than $2$ even when the metric sans outliers is isometrically embeddable into $\ell_2$. We consider bi-criteria approximations. Our main result is a polynomial time algorithm that approximates the outlier set size to within an $O(\log^2 k)$ factor and the distortion to within a constant factor.

The main technical component in our result is an approach for constructing a composition of two given embeddings from subsets of $X$ into $\ell_2$ which inherits the distortions of each to within small multiplicative factors. Specifically, given a low $c_S$ distortion embedding from $S\subset X$ into $\ell_2$ and a high(er) $c_X$ distortion embedding from the entire set $X$ into $\ell_2$, we construct a single embedding that achieves the same  distortion $c_S$ over pairs of points in $S$ and an expansion of at most $O(\log k)\cdot c_X$ over the remaining pairs of points, where $k=|X\setminus S|$. Our composition theorem extends to embeddings into arbitrary $\ell_p$ metrics for $p\ge 1$, and may be of independent interest. While unions of embeddings over disjoint sets have been studied previously, to our knowledge, this is the first work to consider compositions of {\em nested} embeddings.

Abstract:  Graphs arising in statistical problems, signal processing, large networks, combinatorial optimization, and data analysis are often dense, which causes both computational and storage bottlenecks. One way of sparsifying a weighted graph, while sharing the same vertices as the original graph but reducing the number of edges, is through spectral sparsification. We study this problem through the perspective of RandNLA. Specifically, we utilize randomized matrix multiplication to give a clean and simple analysis of how sampling according to edge weights gives a spectral approximation to graph Laplacians, without requiring spectral information. Through the CR−MM algorithm, we attain a simple and computationally efficient sparsifier whose resulting Laplacian estimate is unbiased and of minimum variance. Furthermore, we define a new notion of additive spectral sparsifiers, which has not been considered in the literature.

Abstract