Our paper on scalable quantitative verification for deep neural networks has been accepted to ICSE 2021.
We give a sampling-based approach for quantifying properties for deep neural networks and an attack-agnostic metric called adversarial hardness to capture a model's robustness.
Authors: Teodora Baluta, Zheng Leong Chua, Kuldeep S. Meel and Prateek Saxena
Dec 17, 2020 2:15 PM
We will be presenting three papers at NeurIPS-20.
1. The first paper
focuses on efficient distance approximation in high dimension distributions. We propose an amazingly simple method that can compute L1 distance with rigorous guarantees. Come Poster Session 5
Joint work with A. Bhattacharya, S. Gayen, and N.V. Vinodchandran
2. The second paper
provides the first scalable method to test samplers in practice. Barbarik can now test samplers that sample from log-linear models. If you propose a sampling technique but can't prove its correctness, you can now use Barbarik to check its quality. Just the way we use testing for our software. (Poster
Joint work with S. Chakraborty and Y. Pote
3. The third paper
seeks to tame discrete integration with the boon of dimensionality (Yes, the boon not the curse). We extended our IJCAI-15's work of weighted to unweighted counting (thereby increasing the dimensionality) to handle rational weights. (Poster
Joint work with J.M. Dudek and D. Fried
Dec 5, 2020 2:15 PM
All the five papers from our group were accepted to AAAI 2021.
1. Justicia: A Stochastic SAT Approach to Formally Verify Fairness
Authors: Bishwamittra Ghosh, Debabrota Basu, and Kuldeep S Meel
2. Predicting Forest Fire Using Remote Sensing Data And Machine Learning
Authors: Suwei Yang, Massimo Lupascu, and Kuldeep S Meel
3. Symmetric Component Caching for Model Counting on Structured Instances
Authors: Timothy van Bremen, Vincent Derkinderen, Shubham Sharma, Subhajit Roy, and Kuldeep S Meel
4. Counting Maximal Satisfiable Subsets
Authors: Jaroslav Bendik and Kuldeep S. Meel
5. The Power of Literal Equivalence in Model Counting
Authors: Yong Lai, Kuldeep S Meel, and Roland Yap
Dec 2, 2020 12:00 AM
Our Paper on Model Counting meets F0 Estimation (public version forthcoming!) is accepted to PODS 2021.
Authors: Arnab Bhattacharyya, Kuldeep Meel, A. Pavan and N.V. Vinodchandran
Sep 18, 2020 12:00 AM
We presented three papers at CAV 2020.
1. The first paper builds on our CNF-XOR solving paradigm (BIRD) and as a result, the new versions of ApproxMC and UniGen
are faster than ever.
2. The second paper proposes the first algorithm for approximate MUS counting
3. The third one proposes a data-driven approach for Boolean functional synthesis
, which works at the intersection of constrained sampling, machine learning and automated reasoning.
Jul 22, 2020 12:00 AM
We have released the source code of Manthan.
Jul 13, 2020 12:00 AM
Two Papers accepted to SAT 2020.
1. The first paper shows that the currently known bounds for sparse hashing are too weak to be used for algorithms such as ApproxMC. Authors: Durgesh Agarwal, Bhavishya and Kuldeep S. Meel
2. The second paper proposes a new phase selection strategy for SAT solvers. Improving SAT solvers is just a very very hard job and we are excited about the improvements that our proposal brings to the world of SAT solving. Authors: Arijit Shaw and Kuldeep S. Meel
Apr 25, 2020 12:00 AM
Paper on Sparse Hashing for Approximate Model counting accepted to LICS 2020. Authors: S. Akshay and Kuldeep S. Meel
One of the reviews: “Rarely it is that there is a paper that proves a beautiful new theoretical result, explaining and simplifying previous work, and on top of that shows how it can be used to improve state-of-the-art practical algorithms. The paper “Sparse hashing for scalable approximate model counting: theory and practice” achieves exactly that.”
Apr 10, 2020 12:00 AM
Paper accepted to LPAR-23.
The paper formalizes Induction Models on N extending the classical work of Henkin. Authors: A Dileep, Kuldeep S. Meel and Ammar F. Sabili
Apr 8, 2020 12:00 AM
Three Papers accepted at Computer Aided Verification (CAV) 2020 conference.
1. The first paper proposes a new approach that combines sampling+machine learning+MaxSAT to achieve a significant progress in solving Boolean Functional Synthesis. Authors: Priyanka Golia, Subhajit Roy, and Kuldeep S. Meel
2. The second paper builds on our CNF-XOR solving paradigm (BIRD) and as a result, the new versions of ApproxMC and UniGen are faster than ever. Stay tuned for our releases. Authors: Mate Soos, Stephan Gocht, and Kuldeep S. Meel
3. The third paper proposes the first algorithm for approximate MUS counting. Authors: Jaroslav Bendik and Kuldeep S. Meel
Apr 6, 2020 12:00 AM
Sampling-based approach for quantittive quantitative verification of Deep Neural Nets.
We propose a new attack agnostic metric adversarial hardness to capture the model's robustness: https://arxiv.org/pdf/2002.06864.pdf
Authors: Teodora Baluta, Zheng Leong Chua, Kuldeep S. Meel and Prateek Saxena.
Feb 18, 2020 12:00 AM
Our paper on classification rules in relaxed logical form is accepted in ECAI-2020. Authors: Bishwamittra Ghosh, Dmitry Malioutov, and Kuldeep S. Meel
Jan 15, 2020 12:00 AM