AcademiaEducation
Courses include : (CSC2541) Topics in Machine Learning, Introduction to Causality, (CSC2240) Graphs, Matrices, and Continuous Optimization, (CSC2701) Communication for Computer Scientists, (CSC2541) Advanced Topics in ML: Causal-aware Representation Learning, (CSC2130) Empirical Research Methods in Software Engineering Courses include : (CE695) Stochastic Processes, (CE417) Artificial Intelligence, (CE494) Introduction to Computational Biology, (CE282) Linear Algebra, (CE181) Fundamentals of Probability and Statistics, (CE354) Algorithm Design, (CE415) Theory of Formal Languages and Automata, (MAT034) Differential Equations Publications
Deep generative models often assign higher likelihoods to out-of-distribution data from simpler sources, yet never generate such OOD samples, undermining likelihood-based generative models' reliability. We find that regions with high likelihoods are not generated if they have minimal probability mass, often due to data on low-dimensional manifolds. Using local intrinsic dimension estimation, we introduce an improved out-of-distribution detection method pairing likelihoods and intrinsic dimension estimates from a pre-trained generative model, surpassing existing benchmarks. We introduce a method for determining causal graphs from observational data, leveraging the similarities between multivariate heteroscedastic noise models and affine autoregressive flows. Our approach achieves top results in predicting the true causal structure on genomics benchmarks such as Sachs and SynTReN. Combination therapies are essential for overcoming drug resistance, but the surge in drug combinations has made in Silico simulations crucial. Our method not only identifies synergistic drug pairs but also produces full dose-response matrices, enabling optimal dosage determinations and a deeper understanding of drug interactions. Using attention mechanisms, we predict drug responses on specific cell lines, with state-of-the-art results on the NCI-ALMANAC cancer dataset, outpacing traditional drug fingerprints by leveraging graph neural network embeddings. Physarum Polycephalum, a slime mold, exhibits behavior capable of solving shortest path problems, which has inspired a mathematical model for positive linear programs. This paper introduces an extended family of these dynamics and a new algorithm for positive Semi-Definite Programs (SDP), detailing their foundational principles and their application in discrete optimization, including solutions for MaxCut approximations and clique/chromatic number determinations in perfect graphs. Our studies both theoretically and experimentally highlight the challenges, ensure soundness, and confirm convergence. Paper Reviews
Teaching Assistance
|