Vaibhav Saxena
MSc Applied Computing, Dept of Computer Science, University of Toronto
Phone: +1 (416) 986 - 9091 (Canada)
Email: vaibhav[at]cs[dot]toronto[dot]edu

CV / Resume | Google Scholar | Github

I am a recent Master of Science in Applied Computing graduate from the Department of Computer Science, University of Toronto. I specialize in Deep Reinforcement Learning, Model-based Planning, and Sequence Modeling. I obtainted my Bachelor's degree in Mathematics and Computing from the Indian Institute of Technology Guwahati, in 2018. I am also currently a student researcher at the Vector Institute under the supervision of Prof. Jimmy Ba.

My current research involves learning state-space models for long-horizon planning, and multi-modal representation learning. I have also spent a summer at Microsoft as a Software Engineering intern.
Education
  • Master of Science in Applied Computing (Dept. of Computer Science) at the University of Toronto, Sept 2018 - Dec 2019
    Toronto, Ontario, Canada; GPA: 3.88/4.0

    Courses taken:
    • Machine Learning and Data Mining (CSC2515, Fall 2018) (A+)
    • Deep Reinforcement Learning (CSC2541, Fall 2018) (A+)
    • Computational Linguistics (CSC2501, Fall 2018) (B+)
    • Statistical Learning Theory (CSC2547/STA4273, Winter 2019) (A+)
    • Machine Learning for Health (CSC2541, Winter 2019) (A)
    • Spoken Language Processing (CSC2518, Winter 2019) (A+)

  • Bachelor of Technology in Mathematics and Computing at Indian Institute of Technology Guwahati, July 2014 - June 2018 Guwahati, Assam, India; GPA: 8.97/10

Experiences
  • Machine Learning Engineer at Kindred AI Aug 2020 - Present
    Toronto, Ontario, Canada

  • Student Researcher at the Vector Institute May 2019 - Present
    Toronto, Ontario, Canada

  • Software Engineering Intern at Microsoft May 2017 - July 2017
    India Development Center (IDC), Hyderabad, India

  • Research Intern at Hanyang University, May 2016 - July 2016
    ERICA, Ansan, South Korea

Publications

Clockwork Variational Autoencoders
Vaibhav Saxena, Jimmy Ba, Danijar Hafner
Preprint
[Project Webpage] [arXiv] [Twitter]

Dyna-AIL : Adversarial Imitation Learning by Planning
Vaibhav Saxena, Srinivasan Sivanandan, Pulkit Mathur
Beyond "Tabula Rasa" in Reinforcement Learning Workshop, ICLR 2020
[arXiv]

Academic Projects
  • Evaluating GANs using Efficient Real-World Estimate (CSC2515, Fall 2018)
    [pdf] [Github] Advisor: Prof. Roger Grosse

    In this work, we propose a MCMC sampling method using Metropolis-Hastings algorithm for obtaining samples from pr as learned by the discriminator of a GAN, using samples from the generator. We improve upon an existing rejection sampling approach with respect to a higher acceptance percentage and less samples rejected around the modes. We show how this can be used to evaluate a metric containing a reverse KLD component, rather than adhering to a log-likelihood metric, which seems unfair for models like GANs which intend to minimize the Jenson-Shannon divergence in their loss function.

  • Dyna-AIL: Adversarial Imitation Learning by Planning (CSC2541, Fall 2018)
    [arXiv] Advisor: Prof. Jimmy Ba

    In this work, we propose an end-to-end differentiable adversarial imitation learning algorithm in a Dyna-like framework for switching between model-based planning and model-free learning from expert data. Our results on both discrete and continuous environments show that our approach of using model-based planning along with model-free learning converges to an optimal policy with fewer number of environment interactions in comparison to the state-of-the-art learning methods.

  • fastMRI: Enhancing MRI reconstruction through robust k-space interpolation (CSC2541, Winter 2019)
    [pdf] Advisor: Prof. Marzyeh Ghassemi

    In magnetic resonance imaging (MRI), undersampling the k-space is a widely adopted technique for acceleration. However, this places a trade-off between the acquisition speed and the reconstructed image quality. To address this challenge, we explore several novel machine learning frameworks with the potential of constructing ill-posed MR images, caused by k-space undersampling, to accurate high-quality images. We use two different approaches: k-space imputation using De-noising Autoencoder; and image reconstruction using Generative Adversarial Networks. K-space imputation is a less explored process, and we present our analysis which brings out interesting challenges in this aspect of MRI reconstruction. We also implement an end-to-end model which combines both k-space imputation and image reconstruction to generate sharp MRI images from the blurry ones.

Teaching

  • Teaching Assistant, CSC 373: Algorithm Design, Analysis & Complexity Sep. 2018 - Dec. 2018

Personal stuff

I hail from India, born and brought up, and am bilingual with Hindi and English. In my free time I like to play the piano, and occasionally play tennis.

"May the Force be with all of us."

Last Update: Jan, 3rd, 2019; Template: this/that, ce/cette, das/der, kono/sono and zhege/nage.