Debabrota Basu (@BasuDebabrota) profile picture

Researcher. Writer. Robust, private, & ethical machine learning. Bandits, RL. Faculty @InriaScool in @Inria & @CNRS_HdF Teaching @ENS_ULM/@psl_univ, @univ_lille

France
Joined February 2020
Photo by Debabrota Basu (@BasuDebabrota)Photo by Debabrota Basu (@BasuDebabrota)Photo by Debabrota Basu (@BasuDebabrota)Photo by Debabrota Basu (@BasuDebabrota)Photo by Debabrota Basu (@BasuDebabrota)Photo by Debabrota Basu (@BasuDebabrota)Photo by Debabrota Basu (@BasuDebabrota)Photo by Debabrota Basu (@BasuDebabrota)Photo by Debabrota Basu (@BasuDebabrota)Photo by Debabrota Basu (@BasuDebabrota)
As bandits serve as the cornerstone of decision making with incomplete info & sequential interactions (e.g. recommenders), privacy becomes a pressing concern. We show #DifferentialPrivacy for (some) bandits is free. Is it true for contextual bandits also? #COLT2024 #openproblem
We've derived tight lower and upper bounds for differentially private finite-armed & linear bandits, while we lack the same for contextual bandits. At #COLT2024, @achraf_azize presents open problems in contextual bandits with privacy. @BasuDebabrota @Inria_Lille @RechercheUlille
1
10
413
An interesting line-up on interpretable RL. Congrats to @kohler_hector and organisers to pull off the workshop! 😊
3
196
It's fun to revisit the sanctum sanctorum: how does a brain learn? Today at Convention on Mathematics of #Neuroscience & #AI, @GuillaumeAP presents our work with @AdityaGilra on how to design a bio-plausible learning rule rather than backprop type methods to learn a time series.
2
6
366
Yesterday, @achraf_azize presented @satml_conf the nuances of privacy definitions for bandits, information-theoretic lower bounds, and a wrapper to turn algorithms for linear, contextual, and multi-armed bandits "private". #differential_privacy #bandits 👉arxiv.org/pdf/2309.00557
1/2 To define privacy in bandits, we have to ask what are the input and output of a bandit algorithm? What differs if the adversary is interactive or passive? @achraf_azize & @BasuDebabrota address these in their work openreview.net/forum?id=2366….
3
185
Check arxiv.org/pdf/2309.16563 & openreview.net/forum?id=oGIR… to know more on regret bounds and algorithms for bandits with unbounded corruptions with light-and heavy-tailed rewards underneath. @InriaScool @Inria_Lille #Bandits #regret_bounds #ALT2024 @TmlrOrg
What happens in a bandit problem if epsilon fraction of feedback are arbitrarily corrupt? What are the new lower bounds on the regret? Can we design an optimal algorithm for #Bandits_corrupted_by_nature? We address this question in two parts.
5
188
Great piece by @nellylitvak on Mazur's system of #matheduction. From my little journey of training budding theoreticians, I felt it's fundamental to create a safe space to ask questions, make mistakes, accept them w. no shame, build on lessons & repeat! 👉tinyurl.com/doweteachwhatwep…
1
1
144
Debabrota Basu retweeted
2/2 Can you reconstruct a million $ model just by querying it? Yes, you can replicate BERT's predictions with 90% efficacy only with 1k queries! At 5PM, @BasuDebabrota presents #Marich an information-theoretic unification of black-box model reconstruction: github.com/debabrota-basu/ma…
1
3
158
Join us at @NeurIPSConf for details! Have a great conference ahead 😊 #NeurIPS23
Replying to @NeurIPSConf
@NeurIPSConf, @InriaScool members are going to present 6 papers on #RL, #bandits, and #privacy 🎓We invite the attendees of #NeurIPS2023 to visit & enrich us with your questions! Also, contact Scoolmates at NOLA if you want to know more about works & opportunities in Scool🙏
1
232
As following up the interesting new works #EAAMO23, we would like to thank @ACMEAAMO for recognising our work on collective meritocracy and its impact on college admissions/candidate selection. 🙂 For more, check eaamo2022.eaamo.org/papers/b… #EAAMO22 #FairML @Inria_Lille @InriaScool
Replying to @ACMEAAMO
The work of Thomas K. Buening, @MeiravSegal, @BasuDebabrota, @Annemage4FairAI and Christos Dimitrakakis was awarded in the same category. You can read it here: eaamo2022.eaamo.org/papers/b… Congratulations! 🥳 Get inspired and register for #EAAMO23: cvent.me/lRbw7r (2/2)
1
4
495
Thanks to @IJCAIconf for the stage, & @bishwamittra & @ksmeel for the collaboration. Looking forward to discussions on algorithmic audit of ML/AI as a computational & interdisciplinary problem.🙏 #IJCAI2023 #AIaudit @univ_lille @Inria_Lille @InriaScool @NUSingapore @BenRbg
#Tutorial How do you choose your metric of bias? Debabrota Basu @BasuDebabrota @Inria kicking off the T16: Auditing Bias of Machine Learning Algorithms: Tools and Overview in tandem with Bishwamittra Ghosh @bishwamittra @astar_research #fairerAI #IJCAI2023 #formalmethods
1
8
1,057
1/n Regulating AI platforms has been a rising concern across the globe and @Inria. Today @IJCAIconf, we (with @bishwamittra) present a tutorial on our three-level approach to audit bias/(un)fairness of ML algorithms. @Inria_Lille @InriaScool @NUSComputing
1
2
8
278
2/n With existing works & a demo, we address: 1. How to choose a fairness metric for an application? 2. How to measure bias of an ML algorithm given a fairness metric? 3. How to explain the measured bias as contributions of features? 👉 auditing-fairness-tutorial.g… #AI_audit #FairML
1
1
120
1/n Ever wondered why analysis of differentially private optimisers vary across centralized, federated, and distributed ML? We did. Today @icmlconf we present a unified approach to analyse & design DP optimizers in these settings. 👉arxiv.org/pdf/2302.12559.pdf #ICML2023 @InriaScool
1
4
12
1,413
2/n We use theory of noisy fix-point iteration to (a) unify analysis of DP-SGD, DP-CD, DP-ADMM, (b) design DP-ADMM algorithms for centralised ML, #FedML, and #DistributedML. For more details, visit our poster. 🙏@aurelien_bellet @univ_lille @Inria_Lille
2
126
"And I think to myself What a wonderful world!"
L'image du jour : l'hommage dansé du compagnon d'Agnès Lassalle et de ses amis à l'enseignante, en sortant de la cérémonie d'obsèques ( 🎥@F3euskalherri ) #saintjeandeluz
226
Nos cerisiers sont déjà en fleurs, signe (trop) précoce de l'arrivée du printemps ! #Blossom
2
335
It's been a great week to learn more about the new developments in the AI/ML community, and also presenting our works! #aaai2023
Replying to @InriaScool
@InriaScool members are looking forward to present 6 papers in #AAAI2023 and the co-located workshops. Congratulations to all the authors.🥳Wish the @RealAAAI attendees a great conference ahead!😃🙏
3
253
It's always a proud moment when your student presents the first paper. 😊 @NeurIPSConf Achraf is going to present his work on when privacy meets partial information in sequential decision making. Come over and enrich with your comments. 🙏🏼 #NeurIPS2022 @Inria_Lille
Replying to @InriaScool
2/n Bandits are the basis of modern recommendation systems, which brings forth the concern of data privacy. Today @NeurIPSConf, we present the cost of ensuring data privacy in bandits and a generic wrapper to render your favourite bandit algo differentially private. #dataprivacy
1
12
Joining the ML/AI folks at #NOLA for the week-long festivity of #NeurIPS2022🎄Anyone interested in RL/ bandits, privacy/ fairness/ethical AI, and World Cup Football feel free to connect. 😊 Keep an eye on @InriaScool or ping to know more about PhD, postdoc, and intern positions.
1
7