Debabrota Basu

About Me

Hey, you have now landed in Debabrota Basu's webpage. (Pronounced as De-bo-bro-to)

Presently, I am a tenured faculty (ISFP) at the Scool team (previously called SequeL) of Inria Lille Nord Europe in France. My research aims to construct algorithms and systems for developing efficient, robust, private, and ethical learning machines that solve real-life problems.

Methodologically, my work blends statistics, machine learning, and optimization. In the application side, I am interested in developing theoretically-grounded Responsible AI algorithms for sustainable agro-ecology, medical and pharmaceutical applications, energy-efficient and safe autonomous systems, and algorithmic audits.

Presently, I am teaching postgraduate level courses on privacy, responsible machine learning, and research methods in AI at École normale supérieure-PSL University, Université de Lille, and Centrale Lille. Further details about students and courses are available at the teaching page.

Before arriving in Lille, I have been a postdoctoral researcher at the Data Science and AI Division, Department of Computer Science and Engineering, Chalmers University of Technology. I obtained my PhD in Computer Science from the Department of Computer Science, School of Computing, National University of Singapore. Stéphane Bressan and Pierre Senellart have advised my doctoral journey. I earned a B.E. degree with Honours in Electronics and Telecommunication Engineering from the Department of Electronics and Telecommunication Engineering, Jadavpur University.

Outside academia, I write poetry and plays, air my dissent through articles on scientific and social issues, and have acted in plays and cinema. If any of these provokes your curiosity, have a look at the Arts and Articles page.

News

ABS in ICLR 2024.

We are going to present Augmented Bayesian policy Search (ABS), where we use the action-value functions to conduct efficient exploration with deterministic policies.
Presenting "Pure Exploration in Bandits with Linear Constraints" in AISTATS 2024.

We are going to present how constraints can change the decision space to make pure exploration problem either easier or harder. We present algorithms that can leverage this geometry to yield near-optimal sample complexities.
Honoured to be a member of the ELLIS Society (European Laboratory for Learning and Intelligent Systems) and the Paris unit of ELLIS.
Presentation at AAMAS 2024.

How to adapt your model-based RL algorithm when you have many good models but not the one you want? In response, we will present "Reinforcement Learning in the Wild with Maximum Likelihood-based Model Transfer" at AAMAS 2024.
Talk on "Concentrated Differential Privacy for Bandits" at IEEE SaTML 2024.

Our work on nuances of ensuring approximate and interactive DP for different settings of bandits is accepted in SaTML (IEEE Conference on Secure and Trustworthy Machine Learning), 2024.
CRIMED in ALT 2024.

We are going to present CRIMED, an investigation on impact of unbounded stochastic corruption in bandits, at the Conference of Algorithmic Learning Theory (ALT), 2024.
Presented two papers in NeurIPS 2023.

We presented "On the Complexity of Differentially Private Best-Arm Identification with Fixed Confidence" and "Marich: A Query-efficient Distributionally Equivalent Model Extraction Attack" at NeurIPS, 2023.
Three papers accepted in EWRL 2023.

Our works "Interactive and Concentrated Differential Privacy for Bandits", "Pure Exploration in Bandits with Linear Constraints" and "On the Complexity of Differentially Private Best-Arm Identification with Fixed Confidence" are accepted in EWRL (European Workshop on Reinforcement Learning), 2023.
Our work on unifying differentially private optimisation algorithms for centralised, federated, and distributed learning setups is accepted in ICML 2023.
Our work on explaining sources of bias in outcomes of an ML model is accepted in FAccT 2023.
Three papers are accepted in Privacy-Preserving AI and AI for Agriculture and Food Systems workshops of AAAI 2023.

We are going to present "Marich: A Query-efficient Distributionally-Equivalent Model Extraction Attack using Public Data" and "Rényi Differentially Private Bandits" in PPAI. We will also present farm-gym: a gym-infrastructure inspired by decision making problems in farming that allows deployment of RL algorithms, in AIAFS.
Our paper "Bilinear Exponential Family of MDPs: Frequentist Regret Bound with Tractable Exploration & Planning" is accepted in AAAI 2023.

Bilinear Exponential Family of MDPs provides a generic model encompassing tabular and factored MDPs, and linear dynamical systems. We design a tractable RLSVI-type algorithm, BEF-RLSVI, for Bilinear Exponential Family of MDPs with near-optimal regret bounds.
Our paper "On Meritocracy in Optimal Set Selection" gets the Best Student Paper Award ACM EAAMO 2022.

We all heard that "A group is more than the sum of its parts". Then how can you define the merit of a group? How you can select a class of students who are meritorious together? We study these questions in this paper using the Norwegian college admission data.
Our paper "When Privacy Meets Partial Information" is accepted in NeurIPS 2022.

With Achraf Azize, we have proposed the first lower bound for global differential privacy (DP) in stochastic bandits that holds for any family of distributions. We also propose a generic recipe to turn any index-based bandit algorithm to a its optimal global DP version.
The French National Research Agency (ANR) has awarded the young researcher (JCJC) grant.

D. Basu is one of the recepients of French National Research Agency's (ANR) prestigious young researcher (JCJC) grant of 2022 in "Artificial Intelligence and Data Science". Keep an eye here for open positions if you want to join in this journey.
Three papers accepted in EWRL 2022.

Our works "When Privacy Meets Partial Information", "Bilinear Exponential Family of MDPs: Frequentist Regret Bound with Tractable Exploration and Planning" and "On Bayesian Value Function Distributions" are accepted in EWRL (European Workshop on Reinforcement Learning), 2022.
SENTINEL accepted in UAI 2022.

Our work on SENTINEL: Taming Uncertainty with Ensemble-based Distributional Reinforcement Learning is accepted in UAI (Conference on Uncertainty in Artificial Intelligence), 2022.
Commencement of the Inria-INRAE collaboration on "Reslient Agricultural Decision Making under Environmental Risks".

Delighted to be a part of the collaboration between Scool team of Inria and MIAT team of INRAE to investigate "Resilient Agricultural Decision Making under Environmental Risks": a concerning problem in the face of a shifting global climate.
SAAC accepted in RLDM 2022.

Our work on SAAC: Safe Reinforcement Learning as an Adversarial Game of Actor-Critics is accepted in RLDM (Multidisciplinary Conference on Reinforcement Learning and Decision Making), 2022.
Two papers accepted in AAAI 2022.

We have presented two papers "Procrastinated Tree Search: Black-box Optimization with Delayed, Noisy, and Multi-fidelity Feedback" and "Algorithmic Fairness Verification with Graphical Models" (selected for oral presentation) at AAAI, 22.
Establishment of the associate team, RELIANT, with Kyoto University.

Delighted to be a part of the Inria-Kyoto associate team RELIANT. RELIANT is led by Odalric-Ambrym Maillard (Scool, Inria) and Junya Honda (Kyoto Univ.) with the aim to investigate structured multi-armed bandit problems.
Accepted paper in Proceedings of VLDB 2021.

Our paper on "UDO: Universal Database Optimization using Reinforcement Learning" with Immanuel and Junxiong is published in PVLDB. (Check Immanuel's talk at VLDB'22)
Accepted paper in PharML 2021 at ECML/PKDD.

Collaborating with confidentiality is the future to develop more efficient ML models for safe and effective drug design. In this direction, Shirin Tavara presents our work "Federated Learning of Oligonucleotide Drug Molecule Thermodynamics with Differentially Private ADMM-based SVM" (with myself and Alexander Schilep) at PharML 2021 (Workshop on Machine Learning for Pharma and Healthcare Applications).
Accepted paper in SIGMOD 2021.

Our demonstration paper on "Demonstrating UDO: A Unified Approach for Optimizing Transaction Code, Physical Design, and System Parameters via Reinforcement Learning" with Immanuel Trummer and Junxiong Wang is selected for SIGMOD 2021.
Accepted paper in Journal of Marine Science and Engineering (JMSE).

Our paper on "Deep Learning for Deep Waters: An Expert-in-the-Loop Machine Learning Framework for Marine Sciences" is published in JMSE, February 2021 issue. The paper is accessible at this link.
Accepted paper in AAAI 2021 on Verifying Fairness using Stochastic SAT Solvers.

We (with Bishwamittra Ghosh and Kuldeep S. Meel) have proposed a stochastic SAT based approach Justicia to verify fairness of machine learning models. It comes with both experimental scalability and theoretical error bound unlike the existing formal verifiers of group fairness. The preprint is available at: https://arxiv.org/abs/2009.06516.

Open Positions

Postdoc

Topic: RL under Real-life Constraints: Regrets and Algorithms

We are looking for a postdoctoral researcher willing to join the project to develop theory and algorithms for RL under real-life constraints. The candidate will be supported by a scholarship, and supervised by D. Basu of Scool team at Inria Centre at University of Lille. Tenure of the postdoc candidacy is for two years. The candidate is expected to join in the first quarter of 2023. Get in touch for more details.
PhD and Internship

Contact for available PhD and internship positions. Also, check the job page of Scool for updates!