Sandeep Silwal

Learning-Augmented & Data-Driven Algorithms

• Faster Fundamental Graph Algorithms via Learned Predictions
  Justin Y. Chen, Sandeep Silwal, Ali Vakilian, Fred Zhang
  ICML 2022 [pdf] [story]

  We extend the work of Dinitz et al. (NeurIPS, 2021) and study graph algorithms with the help of learned LP variables. We obtain a better bound for bipartite matching by taking the Hungarian method inspired algorithm of Dinitz et al. and incorporating a a flow perspective. We also give a learned variant of the classic Goldberg's algorithm for shortest-paths with negative edges. Interestingly, the vertex potentials that are often used in shortest-path algorithms are also the variables of an appropriate LP. Lastly, we obtain a host of other learning-based results by utilizing standard (efficient) reductions between graph problems. An interesting development was that for the matching case, the challenge was to utilize a (feasible) set of dual predictions whereas for shortest-paths, the challenge was to round a given set of predictions to first be feasible. After that, optimizing is easy: just run Dijkstra's! Our experiments were also on an interesting dataset: graphs where nodes represent countries and edges are currency exchange rates. By taking the logs of the rates, finding the most efficient way to transfer one currency to another is a cool textbook example of shortest-paths in action.

• Triangle and Four Cycle Counting with Predictions in Graph Stream
  Justin Y. Chen, Talya Eden, Piotr Indyk, Honghao Lin, Shyam Narayanan, Ronitt Rubinfeld, Sandeep Silwal, Tal Wagner, David Woodruff, Michael Zhang
  ICLR 2022 [pdf] [story] [slides] [code]

  Counting triangles in a graph stream is hard: we might accidentally miss 'important' edges in the stream which contribute a lot to the global triangle count. This is where a predictor comes in: if we are given hints on which edges are important, the problem can be mitigated by simply keeping them, leading to improved space bounds. While simple, the observation is quite natural. In practice, there are graph datasets which are slow varying across time (think social network for ex). Theoretically, it turns out many prior algorithms can be reframed in the viewpoint of having access to a predictor.

  See the supplementary material here.

• Learning-Augmented $k$-means Clustering
  Jon Ergun, Zhili Feng, Sandeep Silwal, David Woodruff, Samson Zhou
  Spotlight Presentation at ICLR 2022 [pdf] [story] [slides]

  We consider the problem of $k$-means clustering with the help of a predictor which outputs 'noisy' labels for each point. If the predictor's hints are based on a sufficiently accurate clustering, we can obtain a clustering which can be arbitrarily close to optimum, thereby surpassing known approximation lower bounds (without predictions). Interestingly, it is not sufficient to just blindly follow the predictor's advice and one must post-process the hints. Our algorithms are inspired from robust mean estimation. The fact that we can overcome worst-case NP-hardness via learning is very cool to me and is something that should be explored more.

• Learning-based Support Estimation in Sublinear Time
  Talya Eden, Piotr Indyk, Shyam Narayanan, Ronitt Rubinfeld, Sandeep Silwal, Tal Wagner
  Spotlight Presentation at ICLR 2021 [pdf] [story] [video] [poster]

  We consider the sample complexity of estimating the support size of a discrete distribution drawn from a subset of a possibly large domain $[n]$. Without any advice, the sample complexity is $O(n/\log n)$, barely sublinear. In our model, we assume approximate probabilities of each sample as advice. With this augmentation, the sample complexity drops significantly to $O(n^c)$ for $c < 1$. Our experiments were based on internet search queries for which we could learn approximate underlying probabilities from historical data.

  45 min video
  10 min video

Sublinear and Streaming Algorithms

• The White-Box Adversarial Data Stream Model
  Miklos Ajtai, Vladimir Braverman, T.S. Jayram, Sandeep Silwal, Alec Sun, David P. Woodruff, Samson Zhou
  PODS 2022
  [pdf] [story] [slides]

  Streaming algorithms in the presence of an adaptive adversary has been well studied recently. Most papers are in the so called 'black-box' model where the adversary can only observe outputs of the algorithm to design future inputs. Motivated by real-world adversarial attacks, we consider a "white-box" model where the adversary can now also observe the internal randomness used by the algorithm. Unsurprisingly, many tasks become impossible without storing the entire stream. However, one can still obtain non-trivial algorithms. One example is if we assume the adversary is computationally bounded, then we can use the 'shortest integer solution' (SIS) problem to design streaming algorithms: the SIS matrix is used as a sketching matrix. Even if the adversary knows this matrix, it cannot find integer vectors in the kernel which is a property we exploit. More broadly, we employ collision-resistant hash functions as well as the classic Morris counter to design white-box robust algorithms.

• Adversarial Robustness of Streaming Algorithms through Importance Sampling
  Vladimir Braverman, Avinatan Hassidim, Yossi Matias, Mariano Schain, Sandeep Silwal, Samson Zhou
  Silver Best Paper Award at Adversarial ML Workshop at ICML 2021; Final version at NeurIPS 2021
  [pdf] [story] [video] [slides]

  Adversarial robustness has garnered much recent interest in both the ML and TCS communities. Streaming algorithms are a good intermediate step: they have ML applications while being amenable to theoretical study. The adversary can adaptively design future stream elements after observing the algorithm's response to past inputs. Our simple finding is that existing streaming algorithms, e.g. those based on coresets and sampling based methods, are already robust to the stated model of adversaries. This has wide applications such as in numerical linear algebra and clustering. I believe our experiments are also interesting: they highlight the vulnerability of streaming algorithms implemented in standard and widely used software libraries such as Apache Spark.

• Property Testing of LP-Type Problems
  Rogers Epstein, Sandeep Silwal
  ICALP 2020 [pdf] [story] [video]

  We must determine if a constrained optimization problem is feasible or `far' from feasible while having random access to few constraints of the problem. A cool example is determining if a set of high-dimensional points, which are colored red and blue, are linearly seperable, or if a constant fraction of points need to be removed for the set to be seperable. Here the constraints correspond to having random access to the points. We show that for a natural class of optimization problems, called LP-Type problems, one only needs random access to a number of constraints which is independent of the size of the problem (and only depends on the dimension). The key idea underlying our analysis is the fact that for LP-type problems, which generalize LPs, few constraints already determine feasibility (for ex. consider the 'basis' of a LP). Such ideas were already used in prior works such as Clarkson's classic randomized algorithm for LPs and we port them over to property testing.

• Testing Properties of Multiple Distributions with Few Samples
  Maryam Aliakbarpour, Sandeep Silwal
  ITCS 2020 [pdf] [story] [video]

  We consider a noisy version of distribution testing where each sample can be drawn from a possibly different distribution. While this is hopeless in general, we study natural models of similarity across distributions where one can obtain meaningful results.

  See the first 20 mins. here

Dimensionality Reduction and Its Applications

• Dimensionality Reduction for Wasserstein Barycenter
  Zachary Izzo, Sandeep Silwal, Samson Zhou
  NeurIPS 2021 [pdf] [story] [video]

  We obtain dimensionality reduction bounds for the Wasserstein barycenter problem which go beyond the classical Johnson-Lindenstrauss (JL) lemma. The key idea is to realte the problem to (constrained) $k$-means clustering for which such bounds have been known, e.g., see Makarychev et al. (STOC 2019). Surprisingly, one cannot improve upon the JL lemma if we consider the related problem of optimal transport between two discrete distributions.

• Randomized Dimensionality Reduction for Facility Location and Single-Linkage Clustering
  Shyam Narayanan*, Sandeep Silwal*, Piotr Indyk, Or Zamir
  ICML 2021 [pdf] [story] [video]

  We obtain dimensionality reduction bounds for facility location clustering and the minimum spanning tree problem which go beyond the classical Johnson-Lindenstrauss (JL) lemma. For context, the work of Makarychev et al. (STOC 2019) showed that for $k$-means clustering, one can obtain a dimension bound in terms of $\log k$, rather than $\log n$ where $n$ is the number of points. In the problems we consider, there is no parameter "$k$" to parameterize the dimension bound. Instead we use the doubling dimension, which is a measure of intrinsic dimensionality of the dataset: it is at most $O(\log n)$ in the worst-case (i.e. JL bound) but can also be $O(1)$ regardless of $n$. To obtain our bounds, we exploit 'local' properties of our problems inspired by existing algorithms. The paper has a cool experiment and figure showing that doubling dimension is not just an artifact of analysis and its impact can actually be measured in practice.

  40 min video
  5 min video

• Using Dimensionality Reduction to Optimize t-SNE
  Rikhav Shah, Sandeep Silwal
  OPTML Workshop at NeurIPS 2019 [pdf] [story] [poster]

  t-SNE is a popular visualization tool which maps high dimensional data into two or three dimensions. Unfortunately it is also extremely slow. We propose a somewhat counterintuitive strategy to mitigate this: doing a dimensionality reduction (such as PCA) before applying t-SNE (which itself is dimensionality reduction). Qualitatively this does not seem to make a difference to the t-SNE output while making t-SNE up to an order of magnitude faster.

Broader Algorithmic Works

• Hardness and Algorithms for Robust and Sparse Optimization
  Eric Price, Sandeep Silwal, Samson Zhou
  ICML 2022 [pdf coming soon!] [story]

  Optimization problems where one requires sparsity of variables, such as sparse regression or sparse PCA, are an important and well-studied class of problems. Many of these problems are known to be hard in the worst case, and the best theoretical algorithms have an exponential dependence on $k$, the sparsity parameter. We consider a natural version where we are promised that an optimal solution exists. Given this, we design algorithms for such problems with a reduced exponential dependence of $k/2$, all using the same underlying idea. To illustrate this, suppose we want to solve $Ax=b$ where we are promised a $k$-sparse solution $x$ exists. We first guess over all $k/2$ sparse vectors. We then "match" a fixed guess $y$ to the "correct" other $k/2$ sparse vector using fast nearest neighbor search data structures on the query $b-Ay$. This is the classical algorithmic idea of "meet in the middle." We also study (fine-grained) hardness of various formulations of sparse optimization problems and give reductions between sparse and robust regression, filling a gap in the literature.

• Smoothed Analysis of the Condition Number Under Low-Rank Perturbations
  Rikhav Shah, Sandeep Silwal
  RANDOM 2021 [pdf] [story] [video]

  Most works in smoothed analysis assume dense entry-wise perturbation of the input variables. We take a step back and consider a new model of low-rank perturbations. Our main result is that adding a random rank $k$ matrix to a fixed matrix of rank at-least $n-k$ produces a well-conditioned matrix.

• A Concentration Inequality for the Facility Location Problem
  Sandeep Silwal
  Operations Research Letters, Volume 50 [pdf] [story]

  How well does the objective value of a random instance of the (uniform) facility location problem concentrate? The answer is quite surprising. In the unit square with $n$ uniform points, the interval of concentration is at most $O(n^{1/6})$, an odd exponent. The argument requires Talagrand's concentration inequality which heavily exploits the 'local' structure of the facility location problem.

Miscellaneous

• A note on the universality of ESDs of inhomogeneous random matrices
  Vishesh Jain, Sandeep Silwal
  Latin American Journal of Probability and Mathematical Statistics [pdf]

• Directed Random Geometric Graphs
  Jesse Michel, Ramis Movassagh, Sushruth Reddy, Rikhav Shah, Sandeep Silwal
  Journal of Complex Networks [pdf] [story]

  A novel model of random digraphs for real-world directed networks inspired by the classical random geometric graph model. Interestingly, the in-degree and out-degree distributions follow different laws (power-law vs binomial).