Research Career

With the increase in number of smart devices and the compute power on these devices; there is interest in performing model training on the edge devices. In this research we analyze the current landscape of the edge learning technology and develop new methods to enable model training with reduced resource (memory, compute, energy) footprint.

Selected Publications:

• Sauptik Dhar, Junyao Guo, Jiayi Liu, Samarth Tripathi, Unmesh Kurup, Mohak Shah, “On-Device Machine Learning: An Algorithms and Learning Theory Perspective”, ACM TIOT, 2021 (to appear).

The success of any machine learning (including deep-learning) algorithm depends on careful tuning of its hyperparameters. However hyperparameter tuning is a non-trivial problem and has been a topic of research for several decades. In this research we develop advanced Hyperparameter optimization algorithms which provides significant speed-ups compared to existing state-of-art algorithms. A part of this research also focus on advanced optimization strategies for Deep Learning.
Selected Publications:

• S.Dhar, U.Kurup, M.Shah, “Stabilizing Bi-Level Hyperparameter Optimization using Moreau-Yosida Regularization,” (Poster) ICML 2020 (AutoML Workshop).

• S.Dhar, U.Kurup, M.Shah, “Moreau-Yosida Regularized Bi-Level Hyperparameter Optimization,” KDD 2020 (AutoML Workshop).

The technique of Universum learning provides a formal mechanism for incorporating a priori knowledge about the application domain, in the form of additional Universum samples. Universum learning has been originally introduced for binary classification (by Vapnik) and works specifically well for limited data settings. In this project we formalize Universum Learning to newer learning problems like, regression, multi-class, single-class (anomaly) detection etc.

Selected Publications:

• S. Dhar, V.Cherkassky, M. Shah, “Multiclass Learning from Contradictions”, NeuRIPS 2019.

• S. Dhar, V. Cherkassky, “Development and Evaluation of Cost-Sensitive Universum SVM”, IEEE Transactions on Systems, MAN, and Cybernetics PART B: Cybernetics, vol. 45, no. 4, pp. 806-817, Apr 2015.

• V. Cherkassky, S. Dhar, and W. Dai, “Practical Conditions for Effectiveness of the Universum Learning,” IEEE Transactions on Neural Networks, vol.22, no. 8, pp. 1241-1255, Aug 2011.

Many machine learning applications involve predictive data-analytic modeling using black-box techniques. A common problem in such studies is understanding/interpretation of the estimated models. This project analyzes the current techniques for visualization and interpretation of black-box models and their applicability under VC-theoretic framework.

Selected Publications:

• V. Cherkassky, S. Dhar, “Interpretation of Black-Box Predictive Models”. Measures of Complexity: Festschrift for Alexey Chervonenkis (Editors: V. Vovk , H. Papadopoulos, A. Gammerman), Oct 2015. ( other versions)

• V. Cherkassky, S. Dhar, “Simple Method for Interpretation of High-Dimensional Nonlinear SVM Classification Models”, Proceedings of the 2010 International Conference on Data Mining, July 2010.

With the advent of big-data many data-mining applications involves analyzing huge amount of data (e.g millions of records and or thousands of variables) and derive insights from it. This necessiates research in large-scale data analysis systems and algorithms. In this project we analyze and extend newer optimization algorithms (as scale) on existing big-data frameworks. The outcome of this research will provide tools for building scalable ML algorithms suitable for deriving insights from big-data.

Selected Publications:

• N. Ramakrishnan, S. Dhar, J. Irion, Scalable Graph SLAM for HD Maps, US Patent App.

• S. Dhar, C. Yi, N. Ramakrishnan, and M. Shah, “ADMM based Scalable Machine Learning on Spark”, IEEE Big Data 2015 , Santa Clara, 2015. (Spark Summit Talk)

• Y. Park, S.Dhar, S. Boyd, M. Shah, “Variable Metric Proximal Gradient Method with Diagonal Barzilai-Borwein Stepsize”, NIPS workshop on Optimization for Machine Learning, 2017

As a member of the Data Science Team we are involved in various projects towards deriving insights from the data available in various domains such as., Automotive, Manufacturing, HealthCare, ThermoTechnology etc. A successful project life-cycle spans from requirements gathering to deployment.

Selected Publications:

• M.Ganser, S. Dhar, U. Kurup, C.Cunha, and A.Gacic, “A Data-Driven approach towards Patient Identification for Tele-health Programs”, IEEE Big Data 2015 (Workshop on Mining Big Data to Improve Clinical Effectiveness).

This project describes the data-analytic modeling of the Minnesota soil chemical data produced by the 2001 metro soil survey and by the 2003 state-wide survey. The chemical composition of the soil is characterized by the concentration of many metal and non-metal constituents, resulting in high-dimensional data. This high dimensionality and possible unknown (nonlinear) correlations in the data make it difficult to analyze and interpret using standard statistical techniques. This project applies a machine learning technique, called Self Organizing Map (SOM), to present the high-dimensional soil data in a 2D format suitable for human understanding and interpretation. This SOM representation enables analysis of the soil chemical concentration trends within the metro area and in the state of Minnesota. These trends are important for various Minnesota regulatory agencies concerned with the concentration of polluting chemical elements due to both (a) human activities, i.e., different industrial land usage, and (b) natural geological factors, such as the geomorphic codes and provenance of glacial sediments

Selected Publications:

• S. Dhar, V. Cherkassky, “Statistical Analysis of the Soil Chemical Survey Data”, Report no. Mn/DOT 2010-22, June 2010. ( Technical Report)

• S. Dhar, V. Cherkassky, “Application of SOM to Analysis of Minnesota Soil Survey Data”, International Joint Conference on Neural Networks (IJCNN), Feb 2011.