Tiago H. Falk is a Full Professor at the Institut national de la recherche scientifique, Centre on Energy, Materials, and Telecommunications, University of Quebec, where he directs the Multisensory/multimodal Signal Analysis and Enhancement Lab. He and his team have published over 300 scientific papers on the use of signal processing for improved machine learning applications in real-world settings. Prof. Falk works closely with national and international industry partners to assure their applications operate reliably “in the wild”. He has several patents covering sensor quality measurement and enhancement methods. Several of his tools have been used as benchmarks in IEEE Challenges and have received Best Paper Awards at leading international conferences, including IEEE ICASSP and IEEE SMC. He is Co-Chair of the Technical Committee (TC) on Brain-Machine Interface Systems of the IEEE Systems, Man and Cybernetics (SMC) Society, member of the IEEE Signal Processing Society TC on Audio and Acoustics Signal Processing, member-at-large of the IEEE SMC Society Board of Governors, a founding member of the IEEE Telepresence Initiative, and Academic Chair of the Canadian Medical and Biological Engineering Society. He is co-Editor of the book “Signal Processing and Machine Learning for Biomedical Big Data,” published by CRC Press in 2018.
Conference : How can signal processing benefit AI?
Saturday 7 may 2022, 13h45 - 14h30 — Amphi rouge
There is no doubt that recent innovations in (deep) machine learning have redefined the performance envelope of several applications. But as the old saying goes “when you invent the ship, you also invent the shipwreck.” This is also true of AI. We know that existing systems are (1) sensitive to changes in the distribution of the training data. They are also (2) vulnerable to adversarial attacks, where carefully crafted noise patterns can cause systems to fail drastically, but with very high confidence in their (erroneous) decisions. And (3) since models are data-hungry and require time-consuming hyperparameter optimization steps, they are not environmentally friendly. As emphasized in a 2019 study, training a model with 200 million parameters had an energy consumption and carbon footprint equivalent of 125 round-trip flights between New York and Beijing. Fast forward to 2022, we are talking about models with trillions of parameters, so more sustainable solutions are drastically needed. In this talk, I will showcase innovations in signal processing that are being applied to tackle these three limitations. Successful applications in the speech, healthcare, and human performance monitoring domains will be shown, thus ultimately responding that, yes, signal processing can be used to benefit AI!