Dinesh Daultani

Dinesh Daultani

Visiting Researcher @ Polytechnique Montreal & PhD Candidate @ Tokyo Tech

Background

I am currently a Visiting Researcher at Polytechnique Montreal specializing in designing efficient neural network architectures. Moreover, I am currently pursuing a Ph.D. from the Tokyo Institute of Technology (Tokodai) in the Dept. of Systems and Control Engineering at Okutomi Tanaka Lab. My primary supervisor is Professor Masayuki Tanaka, and I am working closely with Professor Masatoshi Okutomi.

Previously, I worked at Woven Planet Holdings as a Data Scientist, where I was part of the Self-Driving Recognition/Perception AD/ADAS team from 2021 to 2022. At Woven, I work on research problems related to few-shot object detection, scene tags image classification, and road marking recognition. I worked at Rakuten (Rakuten Institute of Technology - R&D Dept) as a Research Scientist for 3.5 years, where I used deep learning, machine learning, and reinforcement learning to solve computer vision, finance, cybersecurity, and e-commerce related research problems. In Rakuten, I worked under the supervision of Bjorn Stenger, Bruno Charron, and Mitsuru Nakazawa. During my master’s, I worked under the supervision of Prof. James Wolf for my master’s project.

Research Interests

My research interests include Computer Vision, Efficient Deep Learning, Neural Architecture Search, and Reinforcement Learning.

Education

  • Ph.D. in Systems and Control Engineering from Tokyo Institute of Technology (2021 - 2024)
    • Research (Computer Vision): Design efficient neural network using techniques such as knowledge distillation, neural architecture search, and evolutionary algorithms for image classification and object detection
  • Master of Science in Information Systems from Illinois State University
  • Bachelor of Engineering in Computer Science from RGPV (CDGI)

Publications

  • ILIAC: Efficient classification of degraded images using knowledge distillation with cutout data augmentation, Proceedings of Image Processing: Algorithms and Systems Conference, at IS&T Electronic Imaging 2023. Link
  • ILIAC+: Efficient image classification of degraded images using knowledge distillation and data augmentation (Submitted in a journal)

Patents

  • Ensembles optimization using genetic algorithms: Dinesh Daultani, Bruno Charron. Information processing apparatus, method for information processing and program thereof. PCT JP2020/001767, filed on 2020/01/20, granted on 2020/11/27 in JP (6801149). Invented a new approach to find the best-optimized combination of weak supervised models by using evolutionary search and tournament selection approach based on weighted metrics
  • Anomaly detection in cybersecurity: Dinesh Daultani, Bruno Charron. Anomaly detection using user probability. PCT JP2020/037004, filed on 2020/09/29, granted on 2022/06/01 in EPO (EP4006760A1). Invented a new approach for anomaly detection based on the change in user probabilities using multi-modal neural networks (CNN & LSTM)
  • Filed 4 other patents on topics related to azimuth angle and screw inspection using drone images.

Visiting Researcher at Polytechnique Montreal

  • Efficient Neural Networks: Design efficient neural network architectures based on binary neural networks and quantization to reduce the matmul operations in neural network under the supervision of Prof. Jean Pierre David.

Data Scientist II at Woven Planet

  • Scene Tags Classification: Design and develop a production model for the classification of scenes for a wide variety of scenarios based on weather, road type, parking, road structures, traffic conditions, intersections, and so on with macro-f1: 0.65 and micro-f1:0.75 for around 100 classes.

  • Few-shot object detection: Design and develop a few-shot object detection method based on CenterNet architecture to improve the performance of the production model on long-tail distribution objects such as ambulances, bikes, and kickboards based on the images of fish-eye lens cameras.

  • Road Marking Recognition: Developed a production model trained on Japanese/English text images to predict paint on streets/roads for self-driving vehicles using the backbone of the CRNN text recognition model with 99% word accuracy and ~0.2 msec latency on a synthetic dataset.

Computer Vision Researcher at Rakuten (RIT)

  • Continuous Taxonomy Learning: Develop a lifelong continual learning system for change in the hierarchy of images over time by utilizing recurrent neural networks, GCN, and lifelong learning methods based on hierarchical classification loss functions while baselining the results on Open Images v6 dataset.

  • Contextual Multi-Armed Bandits: Develop a contextual bandit model to optimize Ad banners on Rakuten e-commerce website for tens of millions of users based on users’ historical purchase data and historical interactions with banners. Write production code to be used in real-time low latency application and performed thorough testing.

  • Anomaly Detection in Security: Develop a user and entity analysis (UEBA) tool using clustering and multi-modal (CNN & LSTM) learning to detect malware and anomalies in tens of thousands of Rakuten employee’s terminals from network text logs.

  • Credit Scoring: Increased the approval rate of credit card merchant customers by 5% by developing a credibility scoring model and segmentation of merchants that led to an increase in revenue of millions of dollars. Lead a team of engineers for the backend development and design different system architecture diagrams.

  • Market Trading: Gained more than 7% ROI / year in daily trading of products such as forex currency pairs, bonds, and indexes by developing end-to-end machine learning models (LSTM & Traditional ML) based on historical market macroeconomics data for assets under management (AUM) of tens of millions of dollars.

Research Assistant at ISU

  • Helped the professor Dr. Xing Fang in reproducing state-of-the-art research papers in computer vision on MNIST & CIFAR-10 datasets, utilizing Theano & Caffe machine learning libraries on Nvidia Titan X GPUs
  • Trained models to optimize the error rate by combining new image transformation ideas

Teaching Assistant at ISU

  • Taught undergrad students Java programming language and basics of computers in the introductory programming course labs
  • Debugged the students’ programming code and helped them to identify & rectify their errors by explaining logical constructs of object-oriented programming language.
    Courses: IT168 and IT150

Additional Roles

  • Technical Interviewer (Woven): Part of the hiring team to screen candidates, review code challenges, and conduct technical interviews focusing on computer vision basics and algorithms/data structures coding.
  • AI Trainer (Rakuten): Delivered training on image classification & object detection using convolutional neural networks and fundamentals of ML in Rakuten for 150+ employees dispersed around different Asia-Pacific locations.
  • RL Study Group (Rakuten): Founder and organizer of Reinforcement Learning study group and paper reading sessions for Rakuten’s machine learning community employees.

Achievements / Awards

  • Recipient of Tokyo Tech Tsubame Scholarship for Doctoral Students from 2023 to 2024.
  • Certificate of Appreciation for my active participation in Intel AI Academy and presenting my research work at PyCon (Python Conference) 2017.
  • Certificate of Honor for designing a website for the annual technological festival held at CDGI (undergrad) in 2013.

Co-Curricular

  • Intel’s Student Ambassador of Artificial Intelligence at Illinois State University in 2016 - 2017.
  • President and Founder of Machine Learning student interest group under ACM/AITP Society at Illinois State University in 2016 - 2017.
  • President of Indian Student Association RSO at Illinois State University in 2016 - 2017.
  • Member of Graduate Student Advisory Council (GSAC) at Illinois State University in 2016 - 2017.

Invited Talks