arXiv subjects

Sahil Singla

Publications and source records attributed to Sahil Singla.

At least 55 records · Page 3Linked to original sources

arXiv2021

Robust Secretary and Prophet Algorithms for Packing Integer Programs

We study the problem of solving Packing Integer Programs (PIPs) in the online setting, where columns in $[0,1]^d$ of the constraint matrix are revealed sequentially, and the goal is to pick a subset of the columns that sum to at most $B$ in each coordinate while maximizing the objective. Excellent results are known in the secretary setting, where the columns are adversarially chosen, but presented in a uniformly random order. However, these existing algorithms are susceptible to adversarial attacks: they try to "learn" characteristics of a good solution, but tend to over-fit to the model, and hence a small number of adversarial corruptions can cause the algorithm to fail. In this paper, we give the first robust algorithms for Packing Integer Programs, specifically in the recently proposed Byzantine Secretary framework. Our techniques are based on a two-level use of online learning, to robustly learn an approximation to the optimal value, and then to use this robust estimate to pick a good solution. These techniques are general and we use them to design robust algorithms for PIPs in the prophet model as well, specifically in the Prophet-with-Augmentations framework. We also improve known results in the Byzantine Secretary framework: we make the non-constructive results algorithmic and improve the existing bounds for single-item and matroid constraints.

cs.DS↗

arXiv2021

Prefix Discrepancy, Smoothed Analysis, and Combinatorial Vector Balancing

A well-known result of Banaszczyk in discrepancy theory concerns the prefix discrepancy problem (also known as the signed series problem): given a sequence of $T$ unit vectors in $\mathbb{R}^d$, find $\pm$ signs for each of them such that the signed sum vector along any prefix has a small $\ell_\infty$-norm? This problem is central to proving upper bounds for the Steinitz problem, and the popular Komlós problem is a special case where one is only concerned with the final signed sum vector instead of all prefixes. Banaszczyk gave an $O(\sqrt{\log d+ \log T})$ bound for the prefix discrepancy problem. We investigate the tightness of Banaszczyk's bound and consider natural generalizations of prefix discrepancy: We first consider a smoothed analysis setting, where a small amount of additive noise perturbs the input vectors. We show an exponential improvement in $T$ compared to Banaszczyk's bound. Using a primal-dual approach and a careful chaining argument, we show that one can achieve a bound of $O(\sqrt{\log d+ \log\!\log T})$ with high probability in the smoothed setting. Moreover, this smoothed analysis bound is the best possible without further improvement on Banaszczyk's bound in the worst case. We also introduce a generalization of the prefix discrepancy problem where the discrepancy constraints correspond to paths on a DAG on $T$ vertices. We show that an analog of Banaszczyk's $O(\sqrt{\log d+ \log T})$ bound continues to hold in this setting for adversarially given unit vectors and that the $\sqrt{\log T}$ factor is unavoidable for DAGs. We also show that the dependence on $T$ cannot be improved significantly in the smoothed case for DAGs. We conclude by exploring a more general notion of vector balancing, which we call combinatorial vector balancing. We obtain near-optimal bounds in this setting, up to poly-logarithmic factors.

cs.DS↗

arXiv2021

Online Discrepancy with Recourse for Vectors and Graphs

The vector-balancing problem is a fundamental problem in discrepancy theory: given T vectors in $[-1,1]^n$, find a signing $σ(a) \in \{\pm 1\}$ of each vector $a$ to minimize the discrepancy $\| \sum_{a} σ(a) \cdot a \|_{\infty}$. This problem has been extensively studied in the static/offline setting. In this paper we initiate its study in the fully-dynamic setting with recourse: the algorithm sees a stream of T insertions and deletions of vectors, and at each time must maintain a low-discrepancy signing, while also minimizing the amortized recourse (the number of times any vector changes its sign) per update. For general vectors, we show algorithms which almost match Spencer's $O(\sqrt{n})$ offline discrepancy bound, with ${O}(n\cdot poly\!\log T)$ amortized recourse per update. The crucial idea is to compute a basic feasible solution to the linear relaxation in a distributed and recursive manner, which helps find a low-discrepancy signing. To bound recourse we argue that only a small part of the instance needs to be re-computed at each update. Since vector balancing has also been greatly studied for sparse vectors, we then give algorithms for low-discrepancy edge orientation, where we dynamically maintain signings for 2-sparse vectors. Alternatively, this can be seen as orienting a dynamic set of edges of an n-vertex graph to minimize the absolute difference between in- and out-degrees at any vertex. We present a deterministic algorithm with $O(poly\!\log n)$ discrepancy and $O(poly\!\log n)$ amortized recourse. The core ideas are to dynamically maintain an expander-decomposition with low recourse and then to show that, as the expanders change over time, a natural local-search algorithm converges quickly (i.e., with low recourse) to a low-discrepancy solution. We also give strong lower bounds for local-search discrepancy minimization algorithms.

cs.DS↗

arXiv2021

Formal Barriers to Simple Algorithms for the Matroid Secretary Problem

Babaioff et al. [BIK2007] introduced the matroid secretary problem in 2007, a natural extension of the classic single-choice secretary problem to matroids, and conjectured that a constant-competitive online algorithm exists. The conjecture still remains open despite substantial partial progress, including constant-competitive algorithms for numerous special cases of matroids, and an $O(\log \log \text{rank})$-competitive algorithm in the general case. Many of these algorithms follow principled frameworks. The limits of these frameworks are previously unstudied, and prior work establishes only that a handful of particular algorithms cannot resolve the matroid secretary conjecture. We initiate the study of impossibility results for frameworks to resolve this conjecture. We establish impossibility results for a natural class of greedy algorithms and for randomized partition algorithms, both of which contain known algorithms that resolve special cases.

cs.DS↗

arXiv2021

Low Curvature Activations Reduce Overfitting in Adversarial Training

Adversarial training is one of the most effective defenses against adversarial attacks. Previous works suggest that overfitting is a dominant phenomenon in adversarial training leading to a large generalization gap between test and train accuracy in neural networks. In this work, we show that the observed generalization gap is closely related to the choice of the activation function. In particular, we show that using activation functions with low (exact or approximate) curvature values has a regularization effect that significantly reduces both the standard and robust generalization gaps in adversarial training. We observe this effect for both differentiable/smooth activations such as SiLU as well as non-differentiable/non-smooth activations such as LeakyReLU. In the latter case, the "approximate" curvature of the activation is low. Finally, we show that for activation functions with low curvature, the double descent phenomenon for adversarially trained models does not occur.

cs.LG↗

arXiv2021

Bag-of-Tasks Scheduling on Related Machines

We consider online scheduling to minimize weighted completion time on related machines, where each job consists of several tasks that can be concurrently executed. A job gets completed when all its component tasks finish. We obtain an $O(K^3 \log^2 K)$-competitive algorithm in the non-clairvoyant setting, where $K$ denotes the number of distinct machine speeds. The analysis is based on dual-fitting on a precedence-constrained LP relaxation that may be of independent interest.

cs.DS↗

arXiv2021

Perceptual Adversarial Robustness: Defense Against Unseen Threat Models

A key challenge in adversarial robustness is the lack of a precise mathematical characterization of human perception, used in the very definition of adversarial attacks that are imperceptible to human eyes. Most current attacks and defenses try to avoid this issue by considering restrictive adversarial threat models such as those bounded by $L_2$ or $L_\infty$ distance, spatial perturbations, etc. However, models that are robust against any of these restrictive threat models are still fragile against other threat models. To resolve this issue, we propose adversarial training against the set of all imperceptible adversarial examples, approximated using deep neural networks. We call this threat model the neural perceptual threat model (NPTM); it includes adversarial examples with a bounded neural perceptual distance (a neural network-based approximation of the true perceptual distance) to natural images. Through an extensive perceptual study, we show that the neural perceptual distance correlates well with human judgements of perceptibility of adversarial examples, validating our threat model. Under the NPTM, we develop novel perceptual adversarial attacks and defenses. Because the NPTM is very broad, we find that Perceptual Adversarial Training (PAT) against a perceptual attack gives robustness against many other types of adversarial attacks. We test PAT on CIFAR-10 and ImageNet-100 against five diverse adversarial attacks. We find that PAT achieves state-of-the-art robustness against the union of these five attacks, more than doubling the accuracy over the next best model, without training against any of them. That is, PAT generalizes well to unforeseen perturbation types. This is vital in sensitive applications where a particular threat model cannot be assumed, and to the best of our knowledge, PAT is the first adversarial training defense with this property.

cs.LG↗

arXiv2021

Fantastic Four: Differentiable Bounds on Singular Values of Convolution Layers

In deep neural networks, the spectral norm of the Jacobian of a layer bounds the factor by which the norm of a signal changes during forward/backward propagation. Spectral norm regularizations have been shown to improve generalization, robustness and optimization of deep learning methods. Existing methods to compute the spectral norm of convolution layers either rely on heuristics that are efficient in computation but lack guarantees or are theoretically-sound but computationally expensive. In this work, we obtain the best of both worlds by deriving {\it four} provable upper bounds on the spectral norm of a standard 2D multi-channel convolution layer. These bounds are differentiable and can be computed efficiently during training with negligible overhead. One of these bounds is in fact the popular heuristic method of Miyato et al. (multiplied by a constant factor depending on filter sizes). Each of these four bounds can achieve the tightest gap depending on convolution filters. Thus, we propose to use the minimum of these four bounds as a tight, differentiable and efficient upper bound on the spectral norm of convolution layers. We show that our spectral bound is an effective regularizer and can be used to bound either the lipschitz constant or curvature values (eigenvalues of the Hessian) of neural networks. Through experiments on MNIST and CIFAR-10, we demonstrate the effectiveness of our spectral bound in improving generalization and provable robustness of deep networks.

cs.LG↗

arXiv2021

Understanding Failures of Deep Networks via Robust Feature Extraction

Traditional evaluation metrics for learned models that report aggregate scores over a test set are insufficient for surfacing important and informative patterns of failure over features and instances. We introduce and study a method aimed at characterizing and explaining failures by identifying visual attributes whose presence or absence results in poor performance. In distinction to previous work that relies upon crowdsourced labels for visual attributes, we leverage the representation of a separate robust model to extract interpretable features and then harness these features to identify failure modes. We further propose a visualization method aimed at enabling humans to understand the meaning encoded in such features and we test the comprehensibility of the features. An evaluation of the methods on the ImageNet dataset demonstrates that: (i) the proposed workflow is effective for discovering important failure modes, (ii) the visualization techniques help humans to understand the extracted features, and (iii) the extracted insights can assist engineers with error analysis and debugging.

cs.CV↗

arXiv2021

Skew Orthogonal Convolutions

Training convolutional neural networks with a Lipschitz constraint under the $l_{2}$ norm is useful for provable adversarial robustness, interpretable gradients, stable training, etc. While 1-Lipschitz networks can be designed by imposing a 1-Lipschitz constraint on each layer, training such networks requires each layer to be gradient norm preserving (GNP) to prevent gradients from vanishing. However, existing GNP convolutions suffer from slow training, lead to significant reduction in accuracy and provide no guarantees on their approximations. In this work, we propose a GNP convolution layer called Skew Orthogonal Convolution (SOC) that uses the following mathematical property: when a matrix is {\it Skew-Symmetric}, its exponential function is an {\it orthogonal} matrix. To use this property, we first construct a convolution filter whose Jacobian is Skew-Symmetric. Then, we use the Taylor series expansion of the Jacobian exponential to construct the SOC layer that is orthogonal. To efficiently implement SOC, we keep a finite number of terms from the Taylor series and provide a provable guarantee on the approximation error. Our experiments on CIFAR-10 and CIFAR-100 show that SOC allows us to train provably Lipschitz, large convolutional neural networks significantly faster than prior works while achieving significant improvements for both standard and certified robust accuracies.

cs.LG↗

arXiv2021

Fairness Through Robustness: Investigating Robustness Disparity in Deep Learning

Deep neural networks (DNNs) are increasingly used in real-world applications (e.g. facial recognition). This has resulted in concerns about the fairness of decisions made by these models. Various notions and measures of fairness have been proposed to ensure that a decision-making system does not disproportionately harm (or benefit) particular subgroups of the population. In this paper, we argue that traditional notions of fairness that are only based on models' outputs are not sufficient when the model is vulnerable to adversarial attacks. We argue that in some cases, it may be easier for an attacker to target a particular subgroup, resulting in a form of \textit{robustness bias}. We show that measuring robustness bias is a challenging task for DNNs and propose two methods to measure this form of bias. We then conduct an empirical study on state-of-the-art neural networks on commonly used real-world datasets such as CIFAR-10, CIFAR-100, Adience, and UTKFace and show that in almost all cases there are subgroups (in some cases based on sensitive attributes like race, gender, etc) which are less robust and are thus at a disadvantage. We argue that this kind of bias arises due to both the data distribution and the highly complex nature of the learned decision boundary in the case of DNNs, thus making mitigation of such biases a non-trivial task. Our results show that robustness bias is an important criterion to consider while auditing real-world systems that rely on DNNs for decision making. Code to reproduce all our results can be found here: \url{https://github.com/nvedant07/Fairness-Through-Robustness}

cs.LG↗

arXiv2020

Online Learning with Vector Costs and Bandits with Knapsacks

We introduce online learning with vector costs (\OLVCp) where in each time step $t \in \{1,\ldots, T\}$, we need to play an action $i \in \{1,\ldots,n\}$ that incurs an unknown vector cost in $[0,1]^{d}$. The goal of the online algorithm is to minimize the $\ell_p$ norm of the sum of its cost vectors. This captures the classical online learning setting for $d=1$, and is interesting for general $d$ because of applications like online scheduling where we want to balance the load between different machines (dimensions). We study \OLVCp in both stochastic and adversarial arrival settings, and give a general procedure to reduce the problem from $d$ dimensions to a single dimension. This allows us to use classical online learning algorithms in both full and bandit feedback models to obtain (near) optimal results. In particular, we obtain a single algorithm (up to the choice of learning rate) that gives sublinear regret for stochastic arrivals and a tight $O(\min\{p, \log d\})$ competitive ratio for adversarial arrivals. The \OLVCp problem also occurs as a natural subproblem when trying to solve the popular Bandits with Knapsacks (\BwK) problem. This connection allows us to use our \OLVCp techniques to obtain (near) optimal results for \BwK in both stochastic and adversarial settings. In particular, we obtain a tight $O(\log d \cdot \log T)$ competitive ratio algorithm for adversarial \BwK, which improves over the $O(d \cdot \log T)$ competitive ratio algorithm of Immorlica et al. [FOCS'19].

cs.LG↗

arXiv2020

Online Carpooling using Expander Decompositions

We consider the online carpooling problem: given $n$ vertices, a sequence of edges arrive over time. When an edge $e_t = (u_t, v_t)$ arrives at time step $t$, the algorithm must orient the edge either as $v_t \rightarrow u_t$ or $u_t \rightarrow v_t$, with the objective of minimizing the maximum discrepancy of any vertex, i.e., the absolute difference between its in-degree and out-degree. Edges correspond to pairs of persons wanting to ride together, and orienting denotes designating the driver. The discrepancy objective then corresponds to every person driving close to their fair share of rides they participate in. In this paper, we design efficient algorithms which can maintain polylog$(n,T)$ maximum discrepancy (w.h.p) over any sequence of $T$ arrivals, when the arriving edges are sampled independently and uniformly from any given graph $G$. This provides the first polylogarithmic bounds for the online (stochastic) carpooling problem. Prior to this work, the best known bounds were $O(\sqrt{n \log n})$-discrepancy for any adversarial sequence of arrivals, or $O(\log\!\log n)$-discrepancy bounds for the stochastic arrivals when $G$ is the complete graph. The technical crux of our paper is in showing that the simple greedy algorithm, which has provably good discrepancy bounds when the arriving edges are drawn uniformly at random from the complete graph, also has polylog discrepancy when $G$ is an expander graph. We then combine this with known expander-decomposition results to design our overall algorithm.

cs.DS↗

arXiv2020

Improved Truthful Mechanisms for Subadditive Combinatorial Auctions: Breaking the Logarithmic Barrier

We present a computationally-efficient truthful mechanism for combinatorial auctions with subadditive bidders that achieves an $O((\log\!\log{m})^3)$-approximation to the maximum welfare in expectation using $O(n)$ demand queries; here $m$ and $n$ are the number of items and bidders, respectively. This breaks the longstanding logarithmic barrier for the problem dating back to the $O(\log{m}\cdot\log\!\log{m})$-approximation mechanism of Dobzinski from 2007. Along the way, we also improve and considerably simplify the state-of-the-art mechanisms for submodular bidders.

cs.GT↗

arXiv2020

Online Discrepancy Minimization for Stochastic Arrivals

In the stochastic online vector balancing problem, vectors $v_1,v_2,\ldots,v_T$ chosen independently from an arbitrary distribution in $\mathbb{R}^n$ arrive one-by-one and must be immediately given a $\pm$ sign. The goal is to keep the norm of the discrepancy vector, i.e., the signed prefix-sum, as small as possible for a given target norm. We consider some of the most well-known problems in discrepancy theory in the above online stochastic setting, and give algorithms that match the known offline bounds up to $\mathsf{polylog}(nT)$ factors. This substantially generalizes and improves upon the previous results of Bansal, Jiang, Singla, and Sinha (STOC' 20). In particular, for the Komlós problem where $\|v_t\|_2\leq 1$ for each $t$, our algorithm achieves $\tilde{O}(1)$ discrepancy with high probability, improving upon the previous $\tilde{O}(n^{3/2})$ bound. For Tusnády's problem of minimizing the discrepancy of axis-aligned boxes, we obtain an $O(\log^{d+4} T)$ bound for arbitrary distribution over points. Previous techniques only worked for product distributions and gave a weaker $O(\log^{2d+1} T)$ bound. We also consider the Banaszczyk setting, where given a symmetric convex body $K$ with Gaussian measure at least $1/2$, our algorithm achieves $\tilde{O}(1)$ discrepancy with respect to the norm given by $K$ for input distributions with sub-exponential tails. Our key idea is to introduce a potential that also enforces constraints on how the discrepancy vector evolves, allowing us to maintain certain anti-concentration properties. For the Banaszczyk setting, we further enhance this potential by combining it with ideas from generic chaining. Finally, we also extend these results to the setting of online multi-color discrepancy.

cs.DS↗

arXiv2020

Second-Order Provable Defenses against Adversarial Attacks

A robustness certificate is the minimum distance of a given input to the decision boundary of the classifier (or its lower bound). For {\it any} input perturbations with a magnitude smaller than the certificate value, the classification output will provably remain unchanged. Exactly computing the robustness certificates for neural networks is difficult since it requires solving a non-convex optimization. In this paper, we provide computationally-efficient robustness certificates for neural networks with differentiable activation functions in two steps. First, we show that if the eigenvalues of the Hessian of the network are bounded, we can compute a robustness certificate in the $l_2$ norm efficiently using convex optimization. Second, we derive a computationally-efficient differentiable upper bound on the curvature of a deep network. We also use the curvature bound as a regularization term during the training of the network to boost its certified robustness. Putting these results together leads to our proposed {\bf C}urvature-based {\bf R}obustness {\bf C}ertificate (CRC) and {\bf C}urvature-based {\bf R}obust {\bf T}raining (CRT). Our numerical results show that CRT leads to significantly higher certified robust accuracy compared to interval-bound propagation (IBP) based training. We achieve certified robust accuracy 69.79\%, 57.78\% and 53.19\% while IBP-based methods achieve 44.96\%, 44.74\% and 44.66\% on 2,3 and 4 layer networks respectively on the MNIST-dataset.

cs.LG↗

arXiv2020

Prophet Inequalities with Linear Correlations and Augmentations

In a classical online decision problem, a decision-maker who is trying to maximize her value inspects a sequence of arriving items to learn their values (drawn from known distributions), and decides when to stop the process by taking the current item. The goal is to prove a "prophet inequality": that she can do approximately as well as a prophet with foreknowledge of all the values. In this work, we investigate this problem when the values are allowed to be correlated. Since non-trivial guarantees are impossible for arbitrary correlations, we consider a natural "linear" correlation structure introduced by Bateni et al. [ESA 2015] as a generalization of the common-base value model of Chawla et al. [GEB 2015]. A key challenge is that threshold-based algorithms, which are commonly used for prophet inequalities, no longer guarantee good performance for linear correlations. We relate this roadblock to another "augmentations" challenge that might be of independent interest: many existing prophet inequality algorithms are not robust to slight increase in the values of the arriving items. We leverage this intuition to prove bounds (matching up to constant factors) that decay gracefully with the amount of correlation of the arriving items. We extend these results to the case of selecting multiple items by designing a new $(1+o(1))$ approximation ratio algorithm that is robust to augmentations.

cs.DS↗

arXiv2020

Online Vector Balancing and Geometric Discrepancy

We consider an online vector balancing question where $T$ vectors, chosen from an arbitrary distribution over $[-1,1]^n$, arrive one-by-one and must be immediately given a $\pm$ sign. The goal is to keep the discrepancy small as possible. A concrete example is the online interval discrepancy problem where T points are sampled uniformly in [0,1], and the goal is to immediately color them $\pm$ such that every sub-interval remains nearly balanced. As random coloring incurs $Ω(T^{1/2})$ discrepancy, while the offline bounds are $Θ(\sqrt{n \log (T/n)})$ for vector balancing and $1$ for interval balancing, a natural question is whether one can (nearly) match the offline bounds in the online setting for these problems. One must utilize the stochasticity as in the worst-case scenario it is known that discrepancy is $Ω(T^{1/2})$ for any online algorithm. Bansal and Spencer recently show an $O(\sqrt{n}\log T)$ bound when each coordinate is independent. When there are dependencies among the coordinates, the problem becomes much more challenging, as evidenced by a recent work of Jiang, Kulkarni, and Singla that gives a non-trivial $O(T^{1/\log\log T})$ bound for online interval discrepancy. Although this beats random coloring, it is still far from the offline bound. In this work, we introduce a new framework for online vector balancing when the input distribution has dependencies across coordinates. This lets us obtain a $poly(n, \log T)$ bound for online vector balancing under arbitrary input distributions, and a $poly(\log T)$ bound for online interval discrepancy. Our framework is powerful enough to capture other well-studied geometric discrepancy problems; e.g., a $poly(\log^d (T))$ bound for the online $d$-dimensional Tusnády's problem. A key new technical ingredient is an {anti-concentration} inequality for sums of pairwise uncorrelated random variables.

cs.DS↗