CONFERENCE/SEMINAR PRESENTATIONS

February 2021

Machine Learning + X seminar (Crunch Group - Applied Math - Brown University)

Video recording link

DEEP EVIDENTIAL CLASSIFICATION/REGRESSION

Deterministic neural networks (NNs) are increasingly being deployed in safety-critical domains, where calibrated, robust, and efficient measures of uncertainty are crucial. Most techniques that provide uncertainty estimates, including Bayesian NNs, rely on sampling multiple NN parameter sets, making multiple predictions, and subsequently computing the statistics of these predictions (such as mean and variance). In other words, prediction uncertainty is inferred through NN parameter uncertainty. Orthogonally to these approaches, recently published works on evidential classification and regression explicitly model uncertainty by placing priors over the categorical (classification) and the Gaussian (regression) likelihood functions. As a result, by training a deterministic NN that outputs the hyperparameters of these priors (instead of the parameters of the likelihood function as usual) a predictor is constructed that outputs both mean response and uncertainty with a single forward pass.

August 2020

Machine Learning + X seminar (Crunch Group - Applied Math - Brown University)

Video recording link

BAYESIAN DEEP LEARNING: A REVIEW

Making accurate and safe predictions using deep learning requires our models to know what they do not know. In other words, predictions should always be accompanied by the model’s confidence in them. To this aim, the Bayesian framework allows the model to incorporate more than one hypothesis for making predictions based on the plausibility of each hypothesis. In this presentation, the Bayesian framework in conjunction with modern deep learning will be discussed. Topics covered are the motivation for incorporating uncertainty into our predictions, Bayesian hierarchical modeling, hyperparameters and hyperpriors, approximate inference techniques and model selection.

April 2020

Johns Hopkins Extreme Materials Institute

PATH INTEGRALS AND SPARSE REPRESENTATIONS IN COMPUTATIONAL STOCHASTIC DYNAMICS


April 2020
Machine Learning + X seminar (Crunch Group - Applied Math - Brown University)

PATH INTEGRALS AND SPARSE REPRESENTATIONS IN COMPUTATIONAL STOCHASTIC DYNAMICS

Addressing current engineering challenges requires collaborative efforts that span across multiple areas of specialization. In this regard, uncertainty quantification represents a significant complement to fields such as computational mechanics and materials science, and has become an essential branch of contemporary engineering research. In engineering dynamics, the task of uncertainty propagation relates to the development of analytical and numerical methodologies for determining response and reliability statistics of complex systems. Specifically, recent advances in stochastic dynamics, ever increasing computational capabilities and advanced experimental setups have contributed to an exceedingly sophisticated mathematical modeling of related systems and excitations, which consequently, demands the development of novel and versatile solution techniques. In this webinar, promising results will be presented based on path integrals, which are considered a potent tool in theoretical physics. By adapting and extending the path integral concept, multidimensional nonlinear dynamical systems subject to various excitations processes can be addressed with satisfactory accuracy. Indicative results pertain to both integer- and fractional-order systems with smooth and non-smooth nonlinearities subject to nonwhite and non-Gaussian excitation processes. Further, it will be demonstrated how regularization and sparse representation concepts are appropriately exploited for enhancing the computational efficiency of the technique. Preliminary results suggest orders of magnitude gain as compared to Monte Carlo simulation schemes.

July 2019 

Engineering Mechanics Institute International Conference, Lyon, FR

A FUNCTIONAL SERIES FORMULATION OF THE WIENER PATH INTEGRAL FOR ACCURATE STOCHASTIC RESPONSE DETERMINATION OF NONLINEAR SYSTEMS


June 2019
Engineering Mechanics Institute Conference, Caltech, Pasadena, USA

ACCURACY-ENHANCED STOCHASTIC RESPONSE DETERMINATION OF NONLINEAR SYSTEMS VIA THE WIENER PATH INTEGRAL TECHNIQUE AND A FUNCTIONAL SERIES EXPANSION APPROACH


July 2018
8th International Conference on Computational Stochastic Mechanics, Paros, GR

NON-STATIONARY JOINT RESPONSE PDF DETERMINATION OF NONLINEAR SYSTEMS VIA THE WIENER PATH INTEGRAL IN CONJUNCTION WITH A WAVELET BASIS


May 2018
Engineering Mechanics Institute Conference, MIT, Cambridge, USA

WIENER PATH INTEGRAL BASED STOCHASTIC RESPONSE DETERMINATION OF STRUCTURAL SYSTEMS: THE NON-WHITE EXCITATION CASE


May 2018
Engineering Mechanics Institute Conference, MIT, Cambridge, USA

WIENER PATH INTEGRAL AND WAVELETS FOR DETERMINING EFFICIENTLY THE NON-STATIONARY JOINT RESPONSE PDF OF NONLINEAR OSCILLATORS


June 2017

Engineering Mechanics Institute Conference, UCSD, San Diego, USA

SPARSE REPRESENTATIONS AND WIENER PATH INTEGRAL FOR EFFICIENT STOCHASTIC RESPONSE DETERMINATION OF MDOF SYSTEMS


 
  • Twitter
  • LinkedIn
  • Instagram

©2021 by Apostolos Psaros