Computer Science Data

HARSIMRAN SINGH

925-***-**** *****************@*****.*** LinkedIn

EDUCATION

University of Utah M.S. in Computer Science GPA: 4.0/4.0 May 2018 Panjab University, India Bachelor of Engineering in Computer Science GPA: 9.0/10.0 June 2016 TECHNICAL SKILLS

Interests Machine Learning, Deep Learning, Data Privacy, Intelligent Transportation Systems Languages R, Python, C, Core Java, MySQL, MongoDB, html, CSS, PHP, Bash Softwares, Packages Tensor

ow, Keras, Spark, Hadoop, GNUPlot, Octave R packages: kernlab, ggplot2, PrivateLR, dplyr, neuralnet, sparkR Python packages: Numpy, Scipy, Pandas, Matplotlib

RELEVANT EXPERIENCE

University of Utah Jan 2017 - Present

Research Assistant Salt Lake City, Utah

Designed a privacy preserving localization approach for mobile users/receivers to locate mobile transmitters.

Perturbing true location of the user in a meaningful way to preserve the privacy of the user. Correspondingly, changing the RSS values using Kernel based approaches, Manifold alignment etc. Also imputing the RSS values at missing places using sparse matrix completion approach.

Built a scalable system to automate crowd-sourcing of data. Android Application on users mobile phone reports data to central controller on request. Dynamic map of current participating users is built and updated. Design Innovation Center & Indian Institute of Technology, Bombay report May 2015 - May 2016 Research Intern Chandigarh, India

First to use barometer sensor present in mobile phones for low power user activity recognition, road tra c congestion (moving, congestion, stuck) in real time with a stuck state accuracy of 97.6%, boarding-deboarding of a vehicle and occupancy of a bus.

Used Decision Trees, K-nearest neighbors, SVM (linear, polynomial, RBF kernel) to perform the above task.

Used Dynamic Time Warping approach to estimate the most recent path traveled by commuter to reach congested area with 90% accuracy. This can be used for mitigation strategies to avoid grave tra c congestion situations.

Further showed how crowdsourced data along with cross-validation can be used to increase the accuracy SELECTED PROJECTS

Deep Learning in Di erential Privacy report poster

. Studied how machine learning & deep learning approaches can potentially leak sensitive information about a dataset.

. Explored and studied di erent models to preserve the privacy of the users like Di erentially Private Stochastic Gradient Descent, Private Aggregation of Teacher Ensembles (PATE) and its generative variant PATE-G.

. Used Rappor to work on small sized datasets and using the idea of domain knowledge increased the accuracy. Privacy Enabled Unsupervised Activity Recognition report poster

. Used clustering to classify user activities into still, walking and vehicle class.

. Tried to map rate of change of altitude values with GPS speed using k-means clustering. Further portrayed how prediction accuracy varies with privacy levels via utility privacy graph. Image Classi cation

. Running and adapting known Neural Networks architectures (CNNs / RNNs) on datasets like MNIST, CIFAR- 10, CIFAR-100 etc. Further optimizing hyperparameters to increase their performance. PUBLICATIONS AND POSTER

. ‘Barosense:Using Barometer for Tra c Congestion Detection and Path Estimation using Crowdsourcing’[under review]

. ‘Using Barometer for Road Tra c Congestion and Path Estimation ’ Utah Data Science Day, January 2017 [Best Poster Award]

. ‘RoadSphygmo:Using Barometer for Tra c Congestion Detection’, COMSNETS 2016 [Acceptance rate: 27.3%]

Contact this candidate