Mahdi Biparva

Summary

I have been modeling, developing, prototyping, and coding deep learning approaches in computer vision for object classification, detection, and segmentation. I have full-stack skills in developing deep learning approaches from the low-level CUDA kernel implementation to the high-level custom layer definition and novel network architecture designs. I have proposed a number of novel approaches to integrate attention in deep learning models for tasks such as object localization, object segmentation, network pruning, and fine-tuning which have led to a number of conference publications and workshop presentations. I have focused on attention in deep learning in my PhD program.

Work/Research Experience

Research Assistant @ Laboratory for Active and attentive Vision (LAAV)

  • York University - Toronto, ON
  • September 2013 to October 2019

Studying visual attention modeling in deep learning is the focus of the PhD program. We proposed a novel Top-Down selection approach that complements the Bottom-Up feedforward feature transformation in convlutional neural networks. Various aspects of the proposed model are studied in different experimental settings such as object classification with localization, semantic segmentation, network compression using pruning, and attentive transfer learning. Full-stack deep learning development is experienced in the PhD programs ranging from the low- level CUDA kernel development, up to the middle-level C++ to Python wrapper coding, and to the high-level custom layer definition in the PyTorch deep learning framework.

OCR Solution Researcher and Developer

  • ApplyBoard - Kitchener, ON
  • September 2018 to November 2018

ApplyBoard is the first online marketplace uniquely designed for international students to apply to the best institutions in Canada and the USA. Our diverse team is fuelled by a passion for culture and innovation, and we recognize education as a right. A contract-based collaboration with ApplyBoard started to develop a deep learning solution for the automation of Optical Character Recognition (OCR) tasks with high accuracy and reliability. The contribution was research, design, development, evaluation of an OCR system. It is capable of input data pre-processing based on the domain requirements, an inference engine trained to predict with %90 label prediction accuracy, and the final post-processing template-matching for the prediction of a sequence of digits and characters.

Computer Vision Research Engineer (Industrial Internship)

  • Huawei Technologies Noah's Ark Lab Markham, ON, Canada
  • May 2018 to September 2018

The Noah's Ark Lab is the AI research centre for Huawei Technologies, located in Hong Kong, Beijing, Shanghai, London, Paris, Toronto, Montreal, Edmonton, etc. As a PhD intern working with Prof. Richard P. Wildes, we developed an action detection framework using the Two-Speed Network inspired by the information processing in human visual cortex. It extends the successful two-stream action recognition approach by introducing expert streams benefiting from two different spatial and temporal speed of information processing. Experimental evaluation for action recognition and temporal action detection reveals the competency over the state-of-the-art. The action recognition and detection networks are implemented from scratch in PyTorch.

Machine Learning Analyst

  • Shoelace Technologies - Toronto, ON
  • September 2016 to February 2017

Shoelace Technologies Inc. is focused on advertisement re-targeting mainly established on Facebook advertisement ecosystem. The part-time collaboration with Shoelace began from descriptive and explanatory data analysis on the data set containing Shopify clients information, Facebook Insight Statistics, client Shoelace profile, and gained benefits of ad re-targeting. The goal was to first analyze the bivariate statistics on the dataset. It was followed by the predictive data analysis phase with the development of a recommendation system to provide the optimal budget investment strategies for different client cases.

Part-time PhD Research Project

  • York University - Toronto, ON
  • 2015 to 2016

A saliency-based object detection system using convlutional neural networks is developed to work on synthetic datasets of person and car categories generated by a 3D simulator. The goal of the prototype is to test the combination of spatio-temporal saliency-based object proposal generation with convolutional neural networks for the task of object classification with localization in computer-generated environments.

Research Assistant @ Centre of Pattern Recognition and Machine Intelligence (CENPARMI)

  • Concordia University - Montreal, QC
  • May 2011 to August 2013

Investigating the application of Gaussian Process classifiers in the handwriting recognition domain for the task of improving the rejection rate metric to increase the systems's reliability. The goal is to reduce human intervention in Optical Character Recognition (OCR) systems by proposing a probabilistic-driven reliable confidence measure so then the number of rejected samples decreases.

Education

PhD in Computer Science

  • York University - Toronto, ON, Canada
  • September 2013 to October 2019
  • Department of Electrical Engineering and Computer Science
  • CGPA: 3.6 out of 4.0

M. Sc. in Computer Science

  • Concordia University - Montreal, QC, Canada
  • January 2011 to August 2013
  • Department of Computer Science and Software Engineering
  • CGPA: 4.2 out of 4.3

B. Eng. in Computer Engineering

  • Tehran Azad University (Central Branch) - Tehran, Iran
  • Sep 2004 – Jan 2010
  • Department of Computer Engineering
  • CGPA: 16.84 out of 20.0

Publications

  • Amir Rosenfeld, Mahdi Biparva, John Tsotsos, Priming Neural Networks, CVPR Workshop on Mutual Benefits of Cognitive and Computer Vision, 2018
  • Mahdi Biparva, John Tsotsos, STNet: Selective Tuning of Convolutional Networks for Object Localization, ICCV Workshop on Mutual Benefits of Cognitive and Computer Vision, 2017