arXiv subjects

Lili Su

Publications and source records attributed to Lili Su.

At least 55 records · Page 3Linked to original sources

arXiv2022

Global Convergence of Federated Learning for Mixed Regression

This paper studies the problem of model training under Federated Learning when clients exhibit cluster structure. We contextualize this problem in mixed regression, where each client has limited local data generated from one of $k$ unknown regression models. We design an algorithm that achieves global convergence from any initialization, and works even when local data volume is highly unbalanced -- there could exist clients that contain $O(1)$ data points only. Our algorithm first runs moment descent on a few anchor clients (each with $\tildeΩ(k)$ data points) to obtain coarse model estimates. Then each client alternately estimates its cluster labels and refines the model estimates based on FedAvg or FedProx. A key innovation in our analysis is a uniform estimate on the clustering errors, which we prove by bounding the VC dimension of general polynomial concept classes based on the theory of algebraic geometry.

cs.LG↗

arXiv2022

A Non-parametric View of FedAvg and FedProx: Beyond Stationary Points

Federated Learning (FL) is a promising decentralized learning framework and has great potentials in privacy preservation and in lowering the computation load at the cloud. Recent work showed that FedAvg and FedProx - the two widely-adopted FL algorithms - fail to reach the stationary points of the global optimization objective even for homogeneous linear regression problems. Further, it is concerned that the common model learned might not generalize well locally at all in the presence of heterogeneity. In this paper, we analyze the convergence and statistical efficiency of FedAvg and FedProx, addressing the above two concerns. Our analysis is based on the standard non-parametric regression in a reproducing kernel Hilbert space (RKHS), and allows for heterogeneous local data distributions and unbalanced local datasets. We prove that the estimation errors, measured in either the empirical norm or the RKHS norm, decay with a rate of 1/t in general and exponentially for finite-rank kernels. In certain heterogeneous settings, these upper bounds also imply that both FedAvg and FedProx achieve the optimal error rate. To further analytically quantify the impact of the heterogeneity at each client, we propose and characterize a novel notion-federation gain, defined as the reduction of the estimation error for a client to join the FL. We discover that when the data heterogeneity is moderate, a client with limited local data can benefit from a common model with a large federation gain. Numerical experiments further corroborate our theoretical findings.

stat.ML↗

arXiv2022

Experimental Design Networks: A Paradigm for Serving Heterogeneous Learners under Networking Constraints

Significant advances in edge computing capabilities enable learning to occur at geographically diverse locations. In general, the training data needed in those learning tasks are not only heterogeneous but also not fully generated locally. In this paper, we propose an experimental design network paradigm, wherein learner nodes train possibly different Bayesian linear regression models via consuming data streams generated by data source nodes over a network. We formulate this problem as a social welfare optimization problem in which the global objective is defined as the sum of experimental design objectives of individual learners, and the decision variables are the data transmission strategies subject to network constraints. We first show that, assuming Poisson data streams, the global objective is a continuous DR-submodular function. We then propose a Frank-Wolfe type algorithm that outputs a solution within a 1-1/e factor from the optimal. Our algorithm contains a novel gradient estimation component which is carefully designed based on Poisson tail bounds and sampling. Finally, we complement our theoretical findings through extensive experiments. Our numerical evaluation shows that the proposed algorithm outperforms several baseline algorithms both in maximizing the global objective and in the quality of the trained models.

cs.NI↗

arXiv2021

The Power of Random Symmetry-Breaking in Nakamoto Consensus

Nakamoto consensus underlies the security of many of the world's largest cryptocurrencies, such as Bitcoin and Ethereum. Common lore is that Nakamoto consensus only achieves consistency and liveness under a regime where the difficulty of its underlying mining puzzle is very high, negatively impacting overall throughput and latency. In this work, we study Nakamoto consensus under a wide range of puzzle difficulties, including very easy puzzles. We first analyze an adversary-free setting and show that, surprisingly, the common prefix of the blockchain grows quickly even with easy puzzles. In a setting with adversaries, we provide a small backwards-compatible change to Nakamoto consensus to achieve consistency and liveness with easy puzzles. Our insight relies on a careful choice of \emph{symmetry-breaking strategy}, which was significantly underestimated in prior work. We introduce a new method -- \emph{coalescing random walks} -- to analyzing the correctness of Nakamoto consensus under the uniformly-at-random symmetry-breaking strategy. This method is more powerful than existing analysis methods that focus on bounding the number of {\it convergence opportunities}.

cs.DC↗

arXiv2019

On Learning Over-parameterized Neural Networks: A Functional Approximation Perspective

We consider training over-parameterized two-layer neural networks with Rectified Linear Unit (ReLU) using gradient descent (GD) method. Inspired by a recent line of work, we study the evolutions of network prediction errors across GD iterations, which can be neatly described in a matrix form. When the network is sufficiently over-parameterized, these matrices individually approximate {\em an} integral operator which is determined by the feature vector distribution $ρ$ only. Consequently, GD method can be viewed as {\em approximately} applying the powers of this integral operator on the underlying/target function $f^*$ that generates the responses/labels. We show that if $f^*$ admits a low-rank approximation with respect to the eigenspaces of this integral operator, then the empirical risk decreases to this low-rank approximation error at a linear rate which is determined by $f^*$ and $ρ$ only, i.e., the rate is independent of the sample size $n$. Furthermore, if $f^*$ has zero low-rank approximation error, then, as long as the width of the neural network is $Ω(n\log n)$, the empirical risk decreases to $Θ(1/\sqrt{n})$. To the best of our knowledge, this is the first result showing the sufficiency of nearly-linear network over-parameterization. We provide an application of our general results to the setting where $ρ$ is the uniform distribution on the spheres and $f^*$ is a polynomial. Throughout this paper, we consider the scenario where the input dimension $d$ is fixed.

cs.LG↗

arXiv2019

Securing Distributed Gradient Descent in High Dimensional Statistical Learning

We consider unreliable distributed learning systems wherein the training data is kept confidential by external workers, and the learner has to interact closely with those workers to train a model. In particular, we assume that there exists a system adversary that can adaptively compromise some workers; the compromised workers deviate from their local designed specifications by sending out arbitrarily malicious messages. We assume in each communication round, up to $q$ out of the $m$ workers suffer Byzantine faults. Each worker keeps a local sample of size $n$ and the total sample size is $N=nm$. We propose a secured variant of the gradient descent method that can tolerate up to a constant fraction of Byzantine workers, i.e., $q/m = O(1)$. Moreover, we show the statistical estimation error of the iterates converges in $O(\log N)$ rounds to $O(\sqrt{q/N} + \sqrt{d/N})$, where $d$ is the model dimension. As long as $q=O(d)$, our proposed algorithm achieves the optimal error rate $O(\sqrt{d/N})$. Our results are obtained under some technical assumptions. Specifically, we assume strongly-convex population risk. Nevertheless, the empirical risk (sample version) is allowed to be non-convex. The core of our method is to robustly aggregate the gradients computed by the workers based on the filtering procedure proposed by Steinhardt et al. On the technical front, deviating from the existing literature on robustly estimating a finite-dimensional mean vector, we establish a {\em uniform} concentration of the sample covariance matrix of gradients, and show that the aggregated gradient, as a function of model parameter, converges uniformly to the true gradient function. To get a near-optimal uniform concentration bound, we develop a new matrix concentration inequality, which might be of independent interest.

cs.DC↗

arXiv2019

Spike-Based Winner-Take-All Computation: Fundamental Limits and Order-Optimal Circuits

Winner-Take-All (WTA) refers to the neural operation that selects a (typically small) group of neurons from a large neuron pool. It is conjectured to underlie many of the brain's fundamental computational abilities. However, not much is known about the robustness of a spike-based WTA network to the inherent randomness of the input spike trains. In this work, we consider a spike-based $k$--WTA model wherein $n$ randomly generated input spike trains compete with each other based on their underlying statistics, and $k$ winners are supposed to be selected. We slot the time evenly with each time slot of length $1\, ms$, and model the $n$ input spike trains as $n$ independent Bernoulli processes. The Bernoulli process is a good approximation of the popular Poisson process but is more biologically relevant as it takes the refractory periods into account. Due to the randomness in the input spike trains, no circuits can guarantee to successfully select the correct winners in finite time. We focus on analytically characterizing the minimal amount of time needed so that a target minimax decision accuracy (success probability) can be reached. We first derive an information-theoretic lower bound on the decision time. We show that to have a (minimax) decision error $\le δ$ (where $δ\in (0,1)$), the computation time of any WTA circuit is at least \[ ((1-δ) \log(k(n -k)+1) -1)T_{\mathcal{R}}, \] where $T_{\mathcal{R}}$ is a difficulty parameter of a WTA task that is independent of $δ$, $n$, and $k$. We then design a simple WTA circuit whose decision time is \[ O( \log\frac{1}δ+\log k(n-k))T_{\mathcal{R}}). \] It turns out that for any fixed $δ\in (0,1)$, this decision time is order-optimal in terms of its scaling in $n$, $k$, and $T_{\mathcal{R}}$.

q-bio.NC↗

arXiv2019

Distributed Learning with Adversarial Agents Under Relaxed Network Condition

This work studies the problem of non-Bayesian learning over multi-agent network when there are some adversarial (faulty) agents in the network. At each time step, each non-faulty agent collects partial information about an unknown state of the world and tries to estimate true state of the world by iteratively sharing information with its neighbors. Existing algorithms in this setting require that all non-faulty agents in the network should be able to achieve consensus via local information exchange. In this work, we present an analysis of a distributed algorithm which does not require the network to achieve consensus. We show that if every non-faulty agent can receive enough information (via iteratively communicating with neighbors) to differentiate the true state of the world from other possible states then it can indeed learn the true state.

cs.DC↗

arXiv2018

Collaboratively Learning the Best Option on Graphs, Using Bounded Local Memory

We consider multi-armed bandit problems in social groups wherein each individual has bounded memory and shares the common goal of learning the best arm/option. We say an individual learns the best option if eventually (as $t\to \infty$) it pulls only the arm with the highest expected reward. While this goal is provably impossible for an isolated individual due to bounded memory, we show that, in social groups, this goal can be achieved easily with the aid of social persuasion (i.e., communication) as long as the communication networks/graphs satisfy some mild conditions. To deal with the interplay between the randomness in the rewards and in the social interaction, we employ the {\em mean-field approximation} method. Considering the possibility that the individuals in the networks may not be exchangeable when the communication networks are not cliques, we go beyond the classic mean-field techniques and apply a refined version of mean-field approximation: (1) Using coupling we show that, if the communication graph is connected and is either regular or has doubly-stochastic degree-weighted adjacency matrix, with probability $\to 1$ as the social group size $N \to \infty$, every individual in the social group learns the best option. (2) If the minimum degree of the graph diverges as $N \to \infty$, over an arbitrary but given finite time horizon, the sample paths describing the opinion evolutions of the individuals are asymptotically independent. In addition, the proportions of the population with different opinions converge to the unique solution of a system of ODEs. In the solution of the obtained ODEs, the proportion of the population holding the correct opinion converges to $1$ exponentially fast in time. Notably, our results hold even if the communication graphs are highly sparse.

cs.DC↗

arXiv2018

On the Convergence Rate of Average Consensus and Distributed Optimization over Unreliable Networks

We consider the problems of reaching average consensus and solving consensus-based optimization over unreliable communication networks wherein packets may be dropped accidentally during transmission. Existing work either assumes that the link failures affect the communication on both directions or that the message senders {\em know exactly}, in each iteration, how many of their outgoing links are functioning properly. In this paper, we consider directed links, and we {\em do not} require each node know its current outgoing degree. First, we propose and characterize the convergence rate of reaching average consensus. Then we apply our robust consensus update to the classical distributed dual averaging method wherein the consensus update is used as the information aggregation primitive. We show that the local iterates converge to a common optimum of the global objective at rate $O(\frac{1}{\sqrt{t}})$, where $t$ is the number of iterations, matching the failure-free performance of the distributed dual averaging method.

cs.DC↗

arXiv2018

Collaboratively Learning the Best Option, Using Bounded Memory

We consider multi-armed bandit problems in social groups wherein each individual has bounded memory and shares the common goal of learning the best arm/option. We say an individual learns the best option if eventually (as $t \to \infty$) it pulls only the arm with the highest average reward. While this goal is provably impossible for an isolated individual, we show that, in social groups, this goal can be achieved easily with the aid of social persuasion, i.e., communication. Specifically, we study the learning dynamics wherein an individual sequentially decides on which arm to pull next based on not only its private reward feedback but also the suggestions provided by randomly chosen peers. Our learning dynamics are hard to analyze via explicit probabilistic calculations due to the stochastic dependency induced by social interaction. Instead, we employ the mean-field approximation method from statistical physics and we show: (1) With probability $\to 1$ as the social group size $N \to \infty $, every individual in the social group learns the best option. (2) Over an arbitrary finite time horizon $[0, T]$, with high probability (in $N$), the fraction of individuals that prefer the best option grows to 1 exponentially fast as $t$ increases ($t\in [0, T]$). A major innovation of our mean-filed analysis is a simple yet powerful technique to deal with absorbing states in the interchange of limits $N \to \infty$ and $t \to \infty $. The mean-field approximation method allows us to approximate the probabilistic sample paths of our learning dynamics by a deterministic and smooth trajectory that corresponds to the unique solution of a well-behaved system of ordinary differential equations (ODEs). Such an approximation is desired because the analysis of a system of ODEs is relatively easier than that of the original stochastic system.

cs.LG↗

arXiv2018

Finite-time Guarantees for Byzantine-Resilient Distributed State Estimation with Noisy Measurements

This work considers resilient, cooperative state estimation in unreliable multi-agent networks. A network of agents aims to collaboratively estimate the value of an unknown vector parameter, while an {\em unknown} subset of agents suffer Byzantine faults. Faulty agents malfunction arbitrarily and may send out {\em highly unstructured} messages to other agents in the network. As opposed to fault-free networks, reaching agreement in the presence of Byzantine faults is far from trivial. In this paper, we propose a computationally-efficient algorithm that is provably robust to Byzantine faults. At each iteration of the algorithm, a good agent (1) performs a gradient descent update based on noisy local measurements, (2) exchanges its update with other agents in its neighborhood, and (3) robustly aggregates the received messages using coordinate-wise trimmed means. Under mild technical assumptions, we establish that good agents learn the true parameter asymptotically in almost sure sense. We further complement our analysis by proving (high probability) {\em finite-time} convergence rate, encapsulating network characteristics.

eess.SY↗

arXiv2017

Distributed Statistical Machine Learning in Adversarial Settings: Byzantine Gradient Descent

We consider the problem of distributed statistical machine learning in adversarial settings, where some unknown and time-varying subset of working machines may be compromised and behave arbitrarily to prevent an accurate model from being learned. This setting captures the potential adversarial attacks faced by Federated Learning -- a modern machine learning paradigm that is proposed by Google researchers and has been intensively studied for ensuring user privacy. Formally, we focus on a distributed system consisting of a parameter server and $m$ working machines. Each working machine keeps $N/m$ data samples, where $N$ is the total number of samples. The goal is to collectively learn the underlying true model parameter of dimension $d$. In classical batch gradient descent methods, the gradients reported to the server by the working machines are aggregated via simple averaging, which is vulnerable to a single Byzantine failure. In this paper, we propose a Byzantine gradient descent method based on the geometric median of means of the gradients. We show that our method can tolerate $q \le (m-1)/2$ Byzantine failures, and the parameter estimate converges in $O(\log N)$ rounds with an estimation error of $\sqrt{d(2q+1)/N}$, hence approaching the optimal error rate $\sqrt{d/N}$ in the centralized and failure-free setting. The total computational complexity of our algorithm is of $O((Nd/m) \log N)$ at each working machine and $O(md + kd \log^3 N)$ at the central server, and the total communication cost is of $O(m d \log N)$. We further provide an application of our general results to the linear regression problem. A key challenge arises in the above problem is that Byzantine failures create arbitrary and unspecified dependency among the iterations and the aggregated gradients. We prove that the aggregated gradient converges uniformly to the true gradient function.

cs.DC↗

arXiv2016

Defending Non-Bayesian Learning against Adversarial Attacks

This paper addresses the problem of non-Bayesian learning over multi-agent networks, where agents repeatedly collect partially informative observations about an unknown state of the world, and try to collaboratively learn the true state. We focus on the impact of the adversarial agents on the performance of consensus-based non-Bayesian learning, where non-faulty agents combine local learning updates with consensus primitives. In particular, we consider the scenario where an unknown subset of agents suffer Byzantine faults -- agents suffering Byzantine faults behave arbitrarily. Two different learning rules are proposed.

cs.DC↗

arXiv2016

Asynchronous Distributed Hypothesis Testing in the Presence of Crash Failures

This paper addresses the problem of distributed hypothesis testing in multi-agent networks, where agents repeatedly collect local observations about an unknown state of the world, and try to collaboratively detect the true state through information exchange. We focus on the impact of failures and asynchrony (two fundamental factors in distributed systems) on the performance of consensus-based non-Bayesian learning. In particular, we consider the scenario where the networked agents may suffer crash faults, and messages delay can be arbitrarily long but finite. We identify the minimal global detectability of the network for non-Bayesian rule to succeed. In addition, we obtain a generalization of a celebrated result by Wolfowitz and Hajnal to submatrices, which might be of independent interest.

cs.DC↗

arXiv2015

Fault-Tolerant Distributed Optimization (Part IV): Constrained Optimization with Arbitrary Directed Networks

We study the problem of constrained distributed optimization in multi-agent networks when some of the computing agents may be faulty. In this problem, the system goal is to have all the non-faulty agents collectively minimize a global objective given by weighted average of local cost functions, each of which is initially known to a non-faulty agent only. In particular, we are interested in the scenario when the computing agents are connected by an arbitrary directed communication network, some of the agents may suffer from crash faults or Byzantine faults, and the estimate of each agent is restricted to lie in a common constraint set. This problem finds its applications in social computing and distributed large-scale machine learning. The fault-tolerant multi-agent optimization problem was first formulated by Su and Vaidya, and is solved when the local functions are defined over the whole real line, and the networks are fully-connected. In this report, we consider arbitrary directed communication networks and focus on the scenario where, local estimates at the non-faulty agents are constrained, and only local communication and minimal memory carried across iterations are allowed. In particular, we generalize our previous results on fully-connected networks and unconstrained optimization to arbitrary directed networks and constrained optimization. As a byproduct, we provide a matrix representation for iterative approximate crash consensus. The matrix representation allows us to characterize the convergence rate for crash iterative consensus.

cs.DC↗

arXiv2015

Fault-Tolerant Multi-Agent Optimization: Part III

We study fault-tolerant distributed optimization of a sum of convex (cost) functions with real-valued scalar input/output in the presence of crash faults or Byzantine faults. In particular, the goal is to optimize a global cost function $\frac{1}{n}\sum_{i\in \mathcal{V}} h_i(x)$, where $\mathcal{V}=\{1, \ldots, n\}$ is the collection of agents, and $h_i(x)$ is agent $i$'s local cost function, which is initially known only to agent $i$. Since the above global cost function cannot be optimized exactly in presence of crash faults or Byzantine faults, we define two weaker versions of the problem for crash faults and Byzantine faults, respectively. When some agents may crash, the goal for the weaker problem is to generate an output that is an optimum of a function formed as $$C(\sum_{i\in \mathcal{N}} h_i(x)+\sum_{i\in \mathcal{F}} α_i h_i(x)),$$ where $\mathcal{N}$ is the set of non-faulty agents, $\mathcal{F}$ is the set of faulty agents (crashed agents), $0\le α_i\le 1$ for each $i\in \mathcal{F}$ and $C$ is a normalization constant such that $C(|\mathcal{N}|+\sum_{i\in \mathcal{F}} α_i)=1$. We present an iterative algorithm in which each agent only needs to perform local computation, and send one message per iteration. When some agents may be Byzantine, the system cannot take full advantage of the data kept by non-faulty agents. The goal for the associated weaker problem is to generate an output that is an optimum of a function formed as $$\sum_{i\in \mathcal{N}}α_i h_i(x),$$ such that $α_i\geq 0$ for each $i\in \mathcal{N}$ and $\sum_{i\in \mathcal{N}}α_i=1$. We present an iterative algorithm, where only local computation is needed and only one message per agent is sent in each iteration, that ensures that at least $|\mathcal{N}|-f$ agents have weights ($α_i$'s) that are lower bounded by $\frac{1}{2(|\mathcal{N}|-f)}$.

cs.DC↗

arXiv2015

Byzantine Multi-Agent Optimization: Part I

We study Byzantine fault-tolerant distributed optimization of a sum of convex (cost) functions with real-valued scalar input/ouput. In particular, the goal is to optimize a global cost function $\frac{1}{|\mathcal{N}|}\sum_{i\in \mathcal{N}} h_i(x)$, where $\mathcal{N}$ is the set of non-faulty agents, and $h_i(x)$ is agent $i$'s local cost function, which is initially known only to agent $i$. In general, when some of the agents may be Byzantine faulty, the above goal is unachievable, because the identity of the faulty agents is not necessarily known to the non-faulty agents, and the faulty agents may behave arbitrarily. Since the above global cost function cannot be optimized exactly in presence of Byzantine agents, we define a weaker version of the problem. The goal for the weaker problem is to generate an output that is an optimum of a function formed as a convex combination of local cost functions of the non-faulty agents. More precisely, for some choice of weights $α_i$ for $i\in \mathcal{N}$ such that $α_i\geq 0$ and $\sum_{i\in \mathcal{N}}α_i=1$, the output must be an optimum of the cost function $\sum_{i\in \mathcal{N}} α_ih_i(x)$. Ideally, we would like $α_i=\frac{1}{|\mathcal{N}|}$ for all $i\in \mathcal{N}$ -- however, this cannot be guaranteed due to the presence of faulty agents. In fact, we show that the maximum achievable number of nonzero weights ($α_i$'s) is $|\mathcal{N}|-f$, where $f$ is the upper bound on the number of Byzantine agents. In addition, we present algorithms that ensure that at least $|\mathcal{N}|-f$ agents have weights that are bounded away from 0. We also propose a low-complexity suboptimal algorithm, which ensures that at least $\lceil \frac{n}{2}\rceil-ϕ$ agents have weights that are bounded away from 0, where $n$ is the total number of agents, and $ϕ$ ($ϕ\le f$) is the actual number of Byzantine agents.

cs.DC↗