Close

Archit Sakhadeo

About Me

I am a graduate student (MSc) at the University of Alberta in the Department of Computing Science. I am supervised by Prof. Adam White and Prof. Alona Fyshe. My current research interests lie in reinforcement learning, representation learning, and real-world reinforcement learning. I am affiliated with the Reinforcement Learning and Artificial Intelligence Lab and the Alberta Machine Intelligence Institute. I am glad to be a part of this incredible research community that has a very positive academic/research culture.


Things I enjoy doing:

• I actively read, write, learn, think, and discuss ideas for contentment.
• I have recently realized that I enjoy teaching and also working remotely from home.
• I enjoy running, and playing squash.
• I recently also enjoyed learning ballroom dance.
• I am extremely fond of Indian, Asian, and Indianized version of Asian cuisines, and also burgers and Döner Kebabs.
• I absolutely love traveling in Goa, Himachal Pradesh, and Uttarkhand in India. I have previously also enjoyed traveling in Europe especially in Luxembourg, Salzburg, Heidelberg, Bruges, and Paris. In Malaysia, my favourite destination is Langkawi.
• I enjoy trekking in the Himalayan foothills, and visiting wildlife national parks.


I come from Thane (India) - the city known for its lakes, Vada pav, and Misal pav.



General Thoughts

I maintain a blog of my essays here.



Academic Thoughts


The Search:

Since centuries, humans have been attempting to understand the principles that make the mind and that lead to intelligence. Understanding what intelligence is, how to simulate it and make it a reality, how to develop intelligence much more capable than what we ever know of has been a long-standing goal of humanity. It is one of the mightiest goals we know of that even the process of attempting it is a humbling and a fruitful exercise. It is one of the most common questions that many of us ponder over at some point in our life and conclude to incomplete answers. This is my small attempt in finding these answers. I share the same curiosity, passion, and enthusiasm that many others who attempt this goal have. I hope that in my lifetime, we complete the missing pieces of this giant puzzle.


An Approach:

Currently, my research interests lie in reinforcement learning, representation learning, and real-world reinforcement learning. I care about building intelligent agents that are grounded in interaction with their environment. Such agents should make use of domain independent, general-purpose learning mechanisms, architectures and algorithms. Such agents should be able to improve their ability to learn with more experience and should be able to adapt to any environment they are placed in, while respecting the inherent constraints on the agents.


I care about reinforcement learning as it is a powerful framework that explains how intelligence can emerge in agents solely on the basis of their interaction with the world around them. I thoroughly enjoy studying and working in reinforcement learning, as it brings together ideas from several areas of knowledge like psychology, neuroscience, and computing science, making it a very rewarding subject to study. Personally, I also think it has ties to philosophy as many parallels can be drawn between life and the reinforcement learning framework.


I care about representation learning because it forms the basis on which learning is dependent. Many components in the reinforcement learning framework rely on good representations. Good representations should lead to better performance, generalization, and no interference/forgetting. With respect to this, I care about constructing the agent state, which is the agent's perception of where it is in the environment. This includes how well the representations summarize the past history, convey the present state, and predict the future. In light of this, I care about general value functions for prediction-based state representation and also discovery/search of good state features. The idea of discovery of representations, hyperparameters, architectures, and learning algorithms with minimal hand-designed components is powerful and general, and is a potential direction to general intelligence. Recently, I have been interested in this study of discovery which is also referred to as learning to learn or meta-learning. Representation learning naturally leads to my interest in function approximation methods like deep learning.


I care about real-world reinforcement learning, especially its application in industrial control and in robotics. Most of the work done today is in the simulated domain as against the real-world where the agents of the future will be operating. Real-world brings many challenges that are important to be addressed but are not explored to the fullest in simulated domains. Today, reinforcement learning is sample inefficient and this characteristic is not very suitable in the real-world where the data rate is much lower than simulated worlds. However, it is important to note that the most gains will also be from its application in the real-world once it is viable.


Accordingly, I have been studying and working in these areas in different capacities.

Education



  University of Alberta

2019 onwards

Master of Science (Thesis), Department of Computing Science

Area of focus: Reinforcement Learning

Supervisors: Prof. Adam White and Prof. Alona Fyshe

Affiliations: RLAI and AMII



  University of Pune

2014 - 2018

Pune Institute of Computer Technology

Bachelor of Engineering, Department of Computer Engineering

Experience

Max Planck Institute for Informatics

    Title: A Common Sense Knowledge Base Framework

    Position: Research Fellow

    Mentors:
  • Dr. Simon Razniewski
  • &
  • Prof. Dr. Gerhard Weikum

  • Ongoing efforts in the direction of building a complete framework of a commonsense knowledge base. Experiments conducted have achieved better salience and coverage than the existing knowledge bases. Such a generic framework would be domain independent and not manually engineered, but would rely on the accuracy of each stage to automatically generate and validate assertions.

Max Planck Institute for Psycholinguistics

    Title: Exploring Linguistic Semantic Interface with Syntactic Processing using Brain activity

    Position: Research Intern

    Collaborator:
  • Sophie Arana, PhD student

  • By conducting MEG experiments on participants reading a stimuli text corpus, the existence of a soft correlation between the P600 event related potential and the absolute difference of semantic associations was proven in order to study how the brain disambiguates prepositional phrase attachments. Corpus-derived co-occurrence frequencies of stimuli words were used as a measure of semantic associations.

Indian Institute of Technology Kanpur

    Title: Automatic Extractive Text Summarizer

    Position: Research Intern

    Mentor:
  • Prof. Nisheeth Srivastava

  • A hybrid approach of using keyword frequencies (statistical approach) with automatically generated entity relationships (semantic approach) was used to combine the strengths and ameliorate the weaknesses of both the approaches, resulting into better summaries. Achieved 16.44% increase in recall and 2.17% increase in Fscore than the next best tested technique. A survey on 94 participants also suggested that our method's summaries were more human-like than the tested methods' summaries.

Tata Institute of Fundamental Research (NCRA-TIFR)

    Title: GMRT Archival Utility for Data Analysis

    Position: Engineering Intern

    Mentors:
  • Prof. Yogesh Wadadekar
  • &
  • Prof. C H Ishwara Chandra
  • Collaborators:
  • Rathin Desai
  • ,
  • Shubhankar Deshpande
  • ,
  • Shadab Shaikh

  • A data processing pipeline was designed to run on a high performance compute cluster to synthesize images from radio interferometric data from the Giant Metrewave Radio Telescope (GMRT). It helped to reduce the synthesis time to under 10-12 hours from 6+ weeks. It has enabled newer cosmological insights and the creation of one of world's largest catalogs for radio astronomy images. Our work was presented at the 30th General Assembly of The International Astronomical Union in Vienna, the 36th Annual Meeting of The Astronomical Society of India, and is accepted at ADASS 2018.

TEDxPICT

    Position: Co-Founder and Curator

    Links:
  • Event page
  • ,
  • Talks

  • TEDxPICT was founded with the intention of promoting social good, thoughtfulness, and discussion. Started as an informal college-level discussion group, it is now an annual city-level event that hosts talks by people who transformed extraordinary ideas into reality. A global platform is provided to the local community as these talks go online. As a curator, my job entailed ideating on the theme and working alongside credited speakers in developing the content of their talk.

Publications

Projects

Create2 Docker

"Hey human, let me charge myself"


Implemented PPO on the Create2 mobile robot to make it dock from anywhere in the designed environment to its charging station learning completely from scratch. The work was an extension of the Benchmarking Reinforcement Learning Algorithms on Real-World Robots paper by Mahmood et al. Experiments were conducted using two techniques - random start position initialization and Curriculum Learning.

CodeReportPresentation

REINFORCE algorithm

"Value Estimate as a baseline? Yes! Discounting rewards? Maybe not."


Implemented the REINFORCE algorithm on the discrete action space and episodic CartPole v1 task from OpenAI Gym. Tested the effect of discounting rewards and subtracting a state dependent, action independent baseline (value estimate) from the returns on the performance of the REINFORCE algorithm.

CodeReport

Travel Guide

"Travel in India! Which places do you wish to visit? How many days would you wish to travel? Voilà! Your travel schedule is ready!"


Implemented using most optimum approach (for smaller number of places) and genetic algorithm (for huge number of places) to solve Traveling Salesman Problem.

View Project

Automatic Precis Generator

"Let's barter! You give me a huge text, I can give you its abstract summary!"


Implemented an automatic extractive text summarizer using keyword frequency and entity relationship graphs.

View Project

Click here for more projects and codes

Contact

<first_name>: archit
<last_name>: sakhadeo

Email: <last_name>@ualberta.ca

<first_name><last_name>@gmail.com


CV available on request