Resume

General Information

Full Name Suraj Tripathi
Contact surajtripathi.iitd [at] gmail [dot] com, surajt [at] cs [dot] cmu [dot] edu
Languages Hindi (native), English (fluent), Punjabi (beginner)

Research Areas

  • My research work has involved working with different modalities such as language, image, and speech. I was able to observe and learn firsthand the intricacies behind different modalities and how different machine learning techniques perceive each of them. This has vastly improved my technical ability and has given me a strong research foundation to further build upon.
  • I believe that there are still vast areas of research left unexplored before machines and humans are indistinguishable in their ability to understand and react to the natural environment. I hope to bridge this gap between man and machine using AI.
  • I am passionate and excited to work in the field of low resource NLP, multi/cross-lingual summarization, human behavior analysis, language generation, and multimodal machine learning.

Education

  • Aug 21 - NOW
    Carnegie Mellon University, Pittsburgh – Master in Language Technologies
    • Advised by Prof. Teruko Mitamura
    • GPA: 4.14 / 4.33
    • PhD level Courses: Advanced Natural Language Processing, Multilingual Natural Language Processing, Multimodal Machine Learning, Question Answering
  • 2015 - 2017
    Indian Institute of Technology, Delhi – Master of Technology in Computer Science
    • Thesis advised by Prof. Jayadeva
    • Thesis: Exploiting Sparsity to attain Faster Run-time Inference and Compressed Deep Neural Network
    • GPA: 8.39 / 10.00
  • 2011 - 2015
    Jamia Millia Islamia, Delhi – Bachelor of Technology in Computer Science
    • Courses: Artificial Intelligence, Data Mining, Parallel & Distributed System, Computer Graphics, Operating System, Compiler Design, etc.
    • GPA: 9.31 / 10.00

Work Experience

  • Aug 21 - now
    Graduate Research Assistant at CMU
    • Advised by Prof. Teruko Mitamura
    • My research project KAIROS involves schema-guided event and entity extraction, temporal event ordering, and cross-media event grounding. My main contribution is towards the event grounding component where I work on finding the best matching between schema events and extracted IE graphs G events.
    • I implemented a matching algorithm that involves aligning extracted events, entities, and relations to a schema. We make use of a transformer-based model to encode event attributes and associated temporal information for matching and instantiation.
    • I also worked on using automatic approaches to predict event saliency scores to extract important events from graphs G that should be present in instantiated schemas. For this, I worked on a GPT-2 based architecture that assigns event saliency score based on coherency loss when that event is removed from the document.
  • Oct 17 - Jul 21
    Lead Software Engineer at Samsung Research Institute, Bangalore, India
    • I worked on building a state-of-the-art Speech Emotion Recognition system with low latency and a low memory footprint. Our objective was to enhance Samsung’s flagship voice assistant Bixby’s response quality not just in content, but also in emotion as studies had proven that dialogue systems that address emotion enhance user satisfaction. Our proposed framework resulted in more than a 4% increase in overall as well as average class accuracy over the existing state-of-the-art methods. In addition, this model contained 62% fewer parameters compared to the benchmark emotion recognition models. This work was published at INTERSPEECH 2018.
    • Improved intent classification and slot tagging model to make it easier for Samsung's voice assistant Bixby to map between human commands to low-level actions to be performed on the mobile device. Awarded Samsung Citizen Award in 2018 and 2019 for outstanding research contributions and technical excellence.
  • Aug 16 - Jun 17
    M.Tech Thesis
    • Deep Neural Networks are both computational and memory intensive, making them difficult to deploy on mobile systems with limited hardware resources. Here, I specifically worked on exploiting the existing redundancies in DNN weights and neural activations in order to maximize compression. I introduced a novel loss function to achieve sparsity by minimizing a convex upper bound on the Vapnik-Chervonenkis (VC) dimension. I also analyzed the effectiveness of our proposed loss function in combination with techniques like quantization and pruning.
  • Mar - Sep 17
    Alpha Researcher at Trexquant Investment LP
    • Involved in analysis of technical and fundamental data for trading statistical arbitrages in global equities. As Global Alpha Researcher, responsible for developing strong risk-adjusted investment portfolios using the latest research from academia.

    • Computer skills

    • ○ Programming languages: C, C++, Python.
    • ○ Data Structures and Algorithms: Knowledge of concepts used in competitive programming and machine learning research.
    • ○ Frameworks: Pytorch, Tensorflow, Keras, NumPy, SciPy.
    • ○ Database Systems: MySQL.

    Other Interests

    • Sports: Volleyball, Badminton, Cricket
    • Hobbies: Traveling, Reading, Music, Movies
© Copyright 2022 Suraj Tripathi.