Sirisha Rambhatla

selected publications

1. 

   ICLR

   🏆 Travel Award

   NOODL: Provable Online Dictionary Learning and Sparse Coding

   Rambhatla, S., Li, X., and Haupt, J.

   International Conference on Learning Representations (ICLR) 2019

   Abs Bib arXiv PDF Code Slides

   We consider the dictionary learning problem, where the aim is to model the givendata as a linear combination of a few columns of a matrix known as adictionary,where the sparse weights forming the linear combination are known ascoeffi-cients. Since the dictionary and coefficients, parameterizing the linear model areunknown, the corresponding optimization is inherently non-convex. This was amajor challenge until recently, when provable algorithms for dictionary learningwere proposed. Yet, these provide guarantees only on the recovery of the dic-tionary, without explicit recovery guarantees on the coefficients. Moreover, anyestimation error in the dictionary adversely impacts the ability to successfully lo-calize and estimate the coefficients. This potentially limits the utility of existingprovable dictionary learning methods in applications where coefficient recoveryis of interest. To this end, we develop NOODL: a simple Neurally plausible alter-nating Optimization-based Online Dictionary Learning algorithm, which recoversboththe dictionary and coefficientsexactlyat a geometric rate, when initializedappropriately. Our algorithm, NOODL, is also scalable and amenable for largescale distributed implementations in neural architectures, by which we mean thatit only involves simple linear and non-linear operations. Finally, we corroboratethese theoretical results via experimental evaluation of the proposed algorithmwith the current state-of-the-art techniques.

   @article{Rambhatla2019, abbr = {ICLR}, title = {{NOODL}: Provable Online Dictionary Learning and Sparse Coding}, author = {Rambhatla, S. and Li, X. and Haupt, J.}, journal = {International Conference on Learning Representations (ICLR)}, year = {2019}, url = {https://openreview.net/forum?id=HJeu43ActQ}, selected = {true}, arxiv = {1902.11261}, slides = {/posters/NOODL_poster_ICLR.pdf}, pdf = {https://openreview.net/pdf?id=HJeu43ActQ}, code = {https://github.com/srambhatla/NOODL}, img = {NOODL.png}, award = {Travel Award} }

2. 

   NeurIPS

   Provable Online CP/PARAFAC Decomposition of a Structured Tensor via Dictionary Learning

   Rambhatla, S., Li, X., and Haupt, J.

   Advances in Neural Information Processing Systems (NeurIPS) 2020

   Abs Bib arXiv PDF Code Slides

   We consider the problem of factorizing a structured 3-way tensor into itsconstituent Canonical Polyadic (CP) factors. This decomposition, whichcan be viewed as a generalization of singular value decomposition (SVD)for tensors, reveals how the tensor dimensions (features) interact with eachother. However, since the factors area prioriunknown, the correspondingoptimization problems are inherently non-convex. The existing guaranteedalgorithms which handle this non-convexity incur an irreducible error (bias),and only apply to cases where all factors have the same structure. To thisend, we develop a provable algorithm for online structured tensor factor-ization, wherein one of the factors obeys some incoherence conditions, andthe others are sparse. Specifically we show that, under some relatively mildconditions on initialization, rank, and sparsity, our algorithm recovers thefactorsexactly(up to scaling and permutation) at a linear rate. Comple-mentary to our theoretical results, our synthetic and real-world data eval-uations showcase superior performance compared to related techniques.

   @article{Rambhatla19Tensor, abbr = {NeurIPS}, title = {Provable Online {CP/PARAFAC} Decomposition of a Structured Tensor via Dictionary Learning }, author = {Rambhatla, S. and Li, X. and Haupt, J.}, journal = {Advances in Neural Information Processing Systems (NeurIPS)}, slides = {/posters/Poster_TensorNOODL.pdf}, img = {TensorNOODL.png}, url = {https://papers.nips.cc/paper/2020/hash/85b42dd8aae56e01379be5736db5b496-Abstract.html}, pdf = {https://papers.nips.cc/paper/2020/file/85b42dd8aae56e01379be5736db5b496-Paper.pdf}, arxiv = {2006.16442}, code = {https://github.com/srambhatla/TensorNOODL}, selected = {true}, year = {2020} }

3. 

   NeurIPS

   How does this interaction affect me? Interpretable attribution for feature interactions

   Tsang, M., Rambhatla, S., and Liu, Y.

   Advances in Neural Information Processing Systems (NeurIPS) 2020

   Abs Bib arXiv PDF Code

   Machine learning transparency calls for interpretable explanations of how inputs relate to predictions. Feature attribution is a way to analyze the impact of features on predictions. Feature interactions are the contextual dependence between features that jointly impact predictions. There are a number of methods that extract feature interactions in prediction models; however, the methods that assign attributions to interactions are either uninterpretable, model-specific, or non-axiomatic. We propose an interaction attribution and detection framework called Archipelago which addresses these problems and is also scalable in real-world settings. Our experiments on standard annotation labels indicate our approach provides significantly more interpretable explanations than comparable methods, which is important for analyzing the impact of interactions on predictions. We also provide accompanying visualizations of our approach that give new insights into deep neural networks.

   @article{tsang2020does, abbr = {NeurIPS}, title = {How does this interaction affect me? Interpretable attribution for feature interactions}, author = {Tsang, M. and Rambhatla, S. and Liu, Y.}, journal = {Advances in Neural Information Processing Systems (NeurIPS)}, arxiv = {2006.10965}, pdf = {https://proceedings.neurips.cc/paper/2020/file/443dec3062d0286986e21dc0631734c9-Paper.pdf}, url = {https://proceedings.neurips.cc/paper/2020/hash/443dec3062d0286986e21dc0631734c9-Abstract.html}, code = {https://github.com/mtsang/archipelago}, img = {archipelago.png}, selected = {true}, year = {2020} }

4. 

   TSP

   A Dictionary-Based Generalization of Robust PCA With Applications to Target Localization in Hyperspectral Imaging

   Rambhatla, S., Li, X., Ren, J., and Haupt, J.

   IEEE Tran. on Signal Processing 2020

   Abs Bib arXiv PDF Blog Code

   We consider the decomposition of a data matrix assumed to be a superposition of a low-rank matrix and a component which is sparse in a known dictionary, using a convex demixing method. We consider two sparsity structures for the sparse factor of the dictionary sparse component, namely entry-wise and column-wise sparsity, and provide a unified analysis, encompassing both undercomplete and the overcomplete dictionary cases, to show that the constituent matrices can be successfully recovered under some relatively mild conditions on incoherence, sparsity, and rank. We leverage these results to localize targets of interest in a hyperspectral (HS) image based on their spectral signature(s) using the a priori known characteristic spectral responses of the target. We corroborate our theoretical results and analyze target localization performance of our approach via experimental evaluations and comparisons to related techniques.

   @article{Rambhatla20LrTheoApp, abbr = {TSP}, author = {{Rambhatla}, S. and {Li}, X. and {Ren}, J. and {Haupt}, J.}, journal = {IEEE Tran. on Signal Processing}, title = {A Dictionary-Based Generalization of Robust PCA With Applications to Target Localization in Hyperspectral Imaging}, year = {2020}, volume = {68}, code = {https://github.com/srambhatla/Dictionary-based-Robust-PCA}, number = {}, pdf = {https://arxiv.org/pdf/1902.08304v3.pdf}, arxiv = {1902.08304v3}, img = {LplusDA_hyper_spectral.gif}, url = {https://arxiv.org/abs/1902.08304}, pages = {1760-75}, selected = {true}, blog = {2017/HyperSpectral/} }