Title: Less is more: Rethinking Few-Shot Learning

Abstract: The statistical supervised learning framework assumes an input-output set with a joint probability distribution that is reliably represented by the training dataset. The learner is then required to output a prediction rule learned from the training dataset's input-output pairs. In this talk, I will present meaningful insights into the information-theoretic asymptotic equipartition property (AEP) (Shannon,1948) in the context of machine learning, and illuminate some of its potential effects on few-shot learning. I will show theoretical guarantees for reliable learning under the AEP, and for the generalization error with respect to the sample size. I will also present experimental results demonstrating the applicability, robustness, and computational efficiency of few-shot learning frameworks for seismic inversion, optical coherence tomography (OCT) speckle suppression and image deblurring. As adaptation of supervised learning models to unseen domains remains a challenging problem, I will briefly discuss domain-awareness in the context of these applications.The results suggest significant potential for improving learning models' sample efficiency,
generalization, and time complexity, that can therefore be leveraged for practical real-time
applications.