Research Experience and Publications

Graduate Research Project, University of Toronto, Fall 2017 - Present

    A particular type of neural networks inspired by techniques from topological data analysis. The preprint can be found at the bottom of this page. (github - under development)

Graduate Research Project, University of Toronto, Supervised by D. Duvenaud, Winter 2017 - Winter 2019

    KF-LAX: Kronecker-factored curvature estimation for control variate optimization in reinforcement learning. (github)

Graduate Research Project, University of Toronto, Supervised by R. Grosse, Fall 2017

    In this project, I and three other graduate students are working on designing attacks on machine learning methods through adversarial examples. Particularly, we are working on training a type of Generative Adversarial Networks (GANs) to produce adversarial examples against a given classification model. Unlike past GAN-based approaches, we are trying to produce adversarial examples in a continuous domain. In our approach, we use a conditional GAN where the generator is conditioned on the unperturbed sample (image) to create an adversarial example while the discriminator is used to distinguish adversarial from non-adversarial examples. (github)

Graduate Research Assistant, University of Toronto, Supervised by M. Brudno, Fall 2017 - Spring 2019

    In this project, I am looking into the problem of unsupervised classification (clustering) of diseases given DNA methylation (DNAm) values for CpG sites specified on people genome. The data we have consists of a set of 200 patients' samples for about 450k DNAm values. As the number of available samples is significantly smaller than the number of variables, the problem becomes particularly challenging due to its high dimensional nature. While in the supervised setting, a relatively well-performing classifier can be trained to the data [1], in an unsupervised setting, training a classifier is particularly challenging, not only because of the dimensionality but also due to limitations of the medical data such as the batch effects. Currently, I am focusing on designing a suitable unsupervised feature selection and learning algorithm for this task.

Undergraduate Research Assistant, Machine Learning Lab, Sharif University of Technology, Supervised by M. Soleymani, Summer 2015 - Summer 2017

    In this lab, I mainly focused on studying the supervised learning problem of classification in Machine Learning. In particular, we have proposed two different approaches to tackle the multi-label classification problem. The first approach was to map both the input data and the labels into a common space where we can define an appropriate similarity distance. The second approach is to perform dimensionality reduction on the labels using non-negative matrix factorization such that the reduced labels can be well approximated from the input data using a linear regression. We obtained promising results with the second approach and I was also able to analyze the algorithm theoretically. We have submitted a summary of this work to IEEE International Conference on Machine Learning and Applications (ICMLA) and the publication is in its proceeding steps.

Undergraduate Research Assistant, Bioinformatics Research Lab, Sharif University of Technology, Supervised by A. Motahari, Summer 2015 - Winter 2017

    In this lab, I performed research on different aspects of genome assembly. In order to join the bioinformatics research group, I had to first prove that I have the necessary knowledge along with the technical skills needed to be successful in this lab. For this purpose, I was asked to devise an approach to distinguish chimeric and non-chimeric contigs that originated from a genome assembler. I implemented my findings, and successful completion was confirmed by testing my implementation on the contigs obtained from the 19th human chromosome. The research I worked on focused on identifying the necessary and sufficient conditions for a set of potentially tagged reads in pooled DNA sequencing to allow correct assembly of each individual genome uniquely. We identified a set of necessary conditions and are tried to prove that these conditions are sufficient. Also, these conditions used to propose limits on the number and the length of reads in order to achieve a reliable assembly of all the pooled DNA sequences.

Research Collaboration with Performance and Dependability Lab, Sharif University of Technology, Supervised by A. Movaghar, Fall 2015 - Summer 2016

    In collaboration with members of Performance and Dependability Lab, I worked on bipartite social networks with a primary research focus of identifying new metrics that could be used to measure centrality of nodes within the network. We defined a probabilistic measure which is essentially the Hellinger distance between the degree distributions of the neighbors of two different nodes in the network. We performed theoretical and experimental analysis to compare our proposed measure with other well-known centrality measures. We have submitted our results to the Springer Journal of Social Network Analysis and Mining (SNAM) and International Workshop on Social Computing (IWSC).


Computer Skills:

Fluent in:   C/C++, Python, MATLAB, R, SQL, Tensorflow, Pytorch, Keras, Kafka, PyMC3, Django (+ web development languages)

Familiar with:   Java, PHP


Publications: