I'm a graduate student in the ECE (Electrical and Computer Engineering) department at UCLA. I am working with
Prof. Cho-Jui Hsieh
on data selection for finetuning small vision models and small LLMs.
Previously, I was a Pre-doctoral Researcher at
Google DeepMind, India where I worked towards improving image understanding, and making image/video
generation more efficient under the guidance of
Dr. Prateek Jain and
Dr. Sujoy Paul . I am broadly interested in computer vision, multimodal representation learning,
and excited about domain adaptation and few shot learning.
I graduated with a Bachelor's degree in Computer Science from
IIIT-Sri City.
Before joining Google, I worked as a Machine Learning Engineer at Tata Consultancy
Services (TCS), Hyderabad where I worked on building ML models with AutoML toolkits
for output explainability.
I interned for a semester at LimeChat
as an AI software developer, where I was involved in designing
contextual chatbots using Level 3 AI.
In the summer of 2020, I worked on unsupervised segmentation of fish in challenging underwater scenarios
under the guidance of Dr. Brejesh Lall at IIT, Delhi.
In my sophomore year, I got an opportunity to explore a little bioinformatics by working on the
problem of prediction of secondary structure of RNA, under the guidance of
Dr. S. Satapathy at Tezpur University.
When I'm not in front of a computer screen, I am mostly playing my guitar and singing. I am also
an avid table tennis player and enjoy reading books during leisure.
Currently reading:
We propose a novel way for efficient inference for image/video generation models.
We introduce a neat way to effectively scale the decoding process with nested model structures and KV caching.
We design a Bayesian optimization based scheme for selecting the best set of hyperparameters (conv layer #filters, layer activation etc)
for a given CNN and use similarity based filtering for choosing redundancies.
Is it an i or an l: Test-time Adaptation of Text Line Recognition Models.
Debapriya Tula, Sujoy Paul, Gagan Madan, Peter Garst, Reeve Ingle, Gaurav Aggarwal.
[Paper] Arxiv Preprint, 2023.
We introduce the problem of test time adaptation for OCR models, given a single image of a writer's handwriting.
We propose a novel confidence based iterative scheme to choose pseudo labels which are used to effectively train
the model at test time.
Offense Detection in Dravidian Languages using Code-Mixing Index based Focal Loss and Cosine
Normalization.
Tula, Debapriya, Shreyas Ms, Viswanatha Reddy, Pranjal
Sahu, Sumanth Doddapaneni, Prathyush
Potluri, Rohan Sukumaran and Parth Patwa.
[Paper]
[Code] SN Computer Science (Journal), 2022.
We introduce a novel code-mixing index (CMI) based focal loss which circumvents code-mixing in
languages and class imbalance for the task of offence detection in Dravidian languages.
Bitions@DravidianLangTech-EACL2021 - Ensemble of Multilingual Language Models with
Pseudo
Labeling for offence Detection in Dravidian Languages.
Debapriya Tula, Prathyush Potluri,
Shreyas MS, Sumanth Doddapaneni, Pranjal Sahu, Rohan Sukumaran, Parth Patwa. Proceedings
of the 1st Workshop on Speech and Language Technologies for Dravidian Languages EACL 2021.
[Paper]
[Code] European Chapter of the Association for
Computational
Linguistics (EACL) Workshop, 2021.
We use a soft voting ensemble of multilingual models, viz. Distil-mBERT and ULMFiT for this
shared task hosted in EACL 2021.
Our solution ranked 1st for the Malayalam dataset and
ranked 4th and 5th
for Tamil and Kannada, respectively.
Estimating RNA Secondary Structure by Maximizing Stacking Regions.
Sen P., Tula D., Ray S.K., Satapathy S.S..
[Paper]
[Code] International Conference on Computer
Communication and Internet
of Things (ICCCIoT 2020).
We try to predict the most stable secondary structure(s) of an RNA sequence using concepts from
Graph Theory to maximise base pairs, leading to minimum entropy structures.
Awarded the best paper at ICCCIoT, 2020.
Incorporation of transition to transversion ratio and nonsense mutations, improves the
estimation of the number of synonymous and non-synonymous sites in codons.
Suvendra K Ray, Ruksana Aziz, Piyali Sen, Pratyush Kumar Beura, Saurav Das, Debapriya Tula, Madhusmita Dash, Nima Dondu Namsa, Ramesh
Chandra Deka, Edward J Feil, Siddhartha Sankar Satapathy.
[Paper]
[Code] DNA Research (Journal), 2022.
Maximize stacking regions to find the most stable secondary structure(s) of RNA using
concepts from graph theory.
Awarded the best paper at ICCCIoT, 2020.
Teaching Experience
Graduate Teaching Assistant (EC M146 - Introduction to Machine Learning) -
(Apr 2025 to Jun 2025)
Assist undergraduates during discussions and office hours.
Graduate Teaching Assistant (LS 30 - Mathematics for Life Scientists) -
(Jan 2025 to Mar 2025)
Lead 3 labs each of 25 undergraduate students for coding assignment problems and math problems for biological modelling problems.
Graded answer sheets of 500+ students during mid-term and end-term exams.
Teaching Assistant (Advanced Data Structures & Algorithms) -
(Sept 2019 to Dec 2019)
Assist 3rd year undergraduates in solving assignment problems during lab sessions. Frame
questions for solving during tutorial sessions and resolve doubts wrt class lectures.
Teaching Assistant (Data Structures & Algorithms) - (Jan 2020
to Apr 2020)
Assist 2nd year undergraduates in solving assignment problems during lab sessions. Conduct
tutorials for doubt clarification.
Projects
Content Based Image Retrieval
Apply Deep Learning based Computer Vision techniques for searching digital images in large
databases.
Gringotts
Provide a vault to securely store secrets like passwords, keys (GPG/SSH), securely transfer
data
among people etc.
[Code]
[Medium]
StackOverFlow API-recommender
Provide an API recommender for Java APIs for questions asked on StackOverflow.
[Code]
Speech Dereverberation
A system to remove reverb(echo) from sound signals by predicting the reverb’s contribution
in
the present signal.
[Code]
Reads For You
A book recommendation system using user-based collaborative filtering.
[Code]