musically-ut
You are using an outdated browser. Please upgrade your browser or activate Google Chrome Frame to improve your experience.
UtkarshU
is not a palindrome
He is a real person.
He solves problems in and using AI/ML.
He learns and lets learn.
He sometimes goes by musically_ut.
I research, code, blog, and make stuff.
Work
Partner at a Stealth startup
Currently working at the intersection of FinTech and AI/ML.
At Reasonal, I am trying to reinvent content management by implementing the algorithms which Behzad Tabibian (CEO, co-founder) and I have developed over the past several years of research.<br>-->
Research and Teaching
You can find my CV here.
Optimizing Global and Local Objectives
I worked with Amita Singh, Jannicke Baalsrud Hauge and Magnus Wiktorsson, on the problem of bringing together local and global stakeholders in city-wide urban logistics with the help of simulations and multi-objective optimization to solve real world problems. We used SUMO with NSGA-II to optimize city-wide emissions as well as noise in certain neighborhoods as a case study in two European cities.
"Optimizing local and global objectives for sustainable mobility in urban areas" ~ Journal of Urban Mobility (2022); Open Access Paper.
-->
Large-scale randomized experiments for more effective memorization
As a follow up to Memorize, I worked with Manuel Gomez-Rodriguez and the creators of Swift Learning App, Christoph Moser and Graham Lancashire, to design a new algorithm for scheduling lessons Select. We ran large scale randomized experiments to verify that the learning indeed was improved by using the ML based instructions.
"Large-scale randomized experiment reveals machine learning helps people learn and remember more effectively" ~ Nature Science of Learning (2021); Open Access Paper.
A #BehindThePaper blog post.
Networks-Learning/spaced-selection
SciPy and Python contributions; NASA Mars rover 2020
My contributions to the sparse matrix API were recognized by making me a co-contributor and co-author to the Nature Methods paper describing SciPy, which coincided with the release of version 1.0 of the library. Also, this along with my contribution to python/cpython, i.e., the Python programming language, also earned me a badge on GitHub for contributing to the Mars 2020 Helicopter Contributor . I always wanted to put something in space 馃
"SciPy 1.0: fundamental algorithms for scientific computing in Python" ~ Nature Methods (2020); Paper.
"Utkarsh Upadhyay contributed code to 2 repositories used in the Mars 2020 Helicopter Mission: python/cpython, and scipy/scipy." ~ GitHub
Learning to Crawl: and other scheduling problems
With R贸bert Busa-Fekete, Wojciech Kot艂owski, D谩vid P谩l, and Bal谩zs Sz枚r茅nyi, I have looked at hte problem of learning to optimally web-crawl pages while simultaneously learning how often they change. Our conclusions about the properties of the learning algorithm and results about learnability of rates of Poisson processes with partial observability apply to many other problems and scenarios as well. We provide the first sub-linear guarantees for such problems and take the first step in the direction of establishing that given some constraints on the optimization problems (e.g. RedQueen, Memorize) which schedule events in continuous time, learning the rates/parameters of the environment while simultaneously optimizing is possible with zero-regret.
"Learning to Crawl" ~ AAAI (2020); Paper.
On the Complexity of Opinions and Online Discussions
With Abir De, Aasish Pappu, and Manuel Gomez-Rodriguez, I have uncovered a connection between complexity of online discussions and the notion of sign-rank of matrices. This allows us to determine the complexity of online discussions just by looking at the pattern of upvotes/downvotes cast by users on others' comments; the key insight is using humans as oracles and by-passing the nuances of sarcasm and humor often present in online comments.
"On Complexity of Opinions and Online Discussions" ~ WSDM (2019); Paper.
Networks-Learning/discussion-complexity
Deep Reinforcement Learning of Marked Temporal Point Processes
With Abir De and Manuel Gomez-Rodriguez, I have developed a deep reinforcement learning algorithm for controlling agents whose actions are performed, and who receives feedback from the environment, at discrete localized points in continuous real time. This is in contract to the classical RL setup where the actions and rewards (feedback) are synchronously given to the agent at discrete points in time.
"Deep Reinforcement Learning of<br>Marked Temporal Point Processes" ~ NeurIPS (2018); Paper.
Networks-Learning/tpprl
3-minute video summary
Memorize: An Online Algorithm for Optimizing Human Learning
With Behzad Tabibian, Abir De, Ali Zarezade, Bernhard Sch枚lkopf and Manuel Gomez-Rodriguez, I have determined the optimal reviewing schedule to keep knowledge fresh in your memory for optimal recall while minimizing effort spent on...