Enhancing Uncertainty Estimation in Neural Networks

Abstract:

Neural networks are often overconfident about their predictions, which undermines their reliability and trustworthiness. In this presentation, I will present our work entitled Error-Driven Uncertainty Aware Training (EUAT), which aims to enhance the ability of neural classifiers to estimate their uncertainty correctly, namely, to be highly uncertain when they output inaccurate predictions and low uncertain when their output is accurate. The EUAT approach operates during the model's training phase by selectively employing two loss functions depending on whether the training examples are correctly or incorrectly predicted by the model. Next, I will address the current limitations of this method and outline the strategies employed in my ongoing research to address these challenges. In particular, I will emphasize the focus on enhancing the efficiency and generalizability of EUAT.

Bio:

Pedro Mendes is a doctoral student in the CMU-Portugal dual-degree PhD program in Software Engineering at Carnegie Mellon University (CMU) and Computer Science and Engineering at Instituto Superior Técnico (IST). He is advised by Prof. Paolo Romano (IST) and Prof. David Garlan (CMU). His research focuses on enhancing the efficiency and trustworthiness of neural networks (NNs). On the efficiency side, he addressed the problem of hyper-parameter tuning, a critical yet expensive step to optimize NN performance. On the trustworthiness front, his work tackles both adversarial robustness and uncertainty estimation. For adversarial robustness, he investigated the challenges and benefits that arise when performing HPT for models that are adversarially trained. For uncertainty estimation, he develops novel methods for uncertainty-aware training across various domains to ensure high uncertainty for incorrect predictions and low uncertainty for correct ones while improving model quality. His broader interests span optimization, machine learning, adversarial training, artificial intelligence, distributed systems, cloud computing, and computer networks.

www.priberam.com