Representation learning for tensor data

Tensors play a significant role in modeling and understanding high-dimensional and complex problems in machine learning, signal processing, and many other areas. For example, many applications involve images, videos, or other types of data samples that are naturally structured as tensors. With the prevalence of big data in many applications, finding low-dimensional representations of data is crucial to keep models reasonably small in order to ensure their efficient and scalable training using large datasets. In this project, I study representation learning methods that exploit the multimodal structure of tensor data to efficiently learn sparse representations of data. More specifically, I work on developing models that best capture the structure in tensor data as well as designing batch and online algorithms that can deal with big and high-dimensional datasets as well as streaming settings.

Matrix Completion for Recommender Systems 2017-2018

I started working on this problem during my internship at Technicolor AI Lab. I implemented different matrix and tensor completion algorithms on real-world datasets during my time at technicolor. Moreover, I was particularly interested in the optimization landscape of the nonconvex inductive matrix completion problem (a matrix completion problem with side information in form of features of users and items). My work builds upon the many works in the recent yeras on the optimization landscape of nonconvex problems such as matrix factorization, deep learning, etc. that are motivated by success of first order optimization algorithms in solving such prolems.

Communication-Efficient Decentralized Optimization 2017-2018

During my masters studies at Rutgers I worked on a coordinate-wise variant of Decentralized Gradient Descent (DGD) that considerably reduces the amount of communication per iteration by only sharing the value of one coordinate at a time rather than the entire decision vector. This scheme is especially useful when dealing with very high dimensional data.

Differentially Private Active Learning 2017-2018

Differential privacy makes it possible to collect and share aggregate information about users, while preserving the privacy of individual users. This makes it a suitable privacy notion for machine learning where presence of an individual data sample should not considerably affect the model learned using the entire dataset. In this project I developed a differentially private active learning algorithm for anomaly detection in streaming data settings.

Deep Learning for Face Recognition

As a course project for the graduate course "Deep Learning", we were tasked with fine-tuning and evaluating the performance of Alexnet and VGG-16 pre-trained networks on the Labeled Faces in the Wild (LFW) dataset. This project was a great opportunity to get hands-on experience working with convolutional neural nets (CNNs) and become familiar with TensorFlow.

Sentiment Analysis of IMDb movie reviews

This project exposed me to various approaches that exploit semantic relationship between words to create meaningful representations for words. We first tried to find the most informative words using information retrieval approaches such as mutual information. Then, we tried different vector space models (VSMs) to learn vector representations of the reviews. We tried bag of words, tf-idf, and neural network-based methods word2vec and doc2vec. We tried different classification methods such as logistic regression, SVM, and random forest for sentiment classification. We achieved best results by merging two representation approaches. That is, for each review, we stacked the feature vector from the tf­-idf approach with the feature vector from the Doc2vec approach to train a random forest classifier. This project gave me experience working with python packages like numpy, scipy, and scikit-learn.