This is an online seminar.
Registration is required.
【Speaker】Christopher Johannes Anders
Technische Universität Berlin, Machine Learning Group, Doctoral Researcher
【Title】
Debugging Learning Algorithms: Understanding and Correcting Machine Learning Models
【Abstract】
Despite the immense prediction capabilities of deep neural networks demonstrated in various domains, their complexity denies us any deeper understanding of their prediction strategies. Despite this, they are
often trained on incomprehensibly large bodies of data with high potential for undetected biases.
Training on such data may thus result in the undetected learning of biased decision strategies. Common model performance metrics fail to uncover such bugs, requiring a deeper understanding of the model itself.
In this talk, we present our work on debugging deep neural networks based on approaches from explainable artificial intelligence (XAI). We demonstrate that the importance of individual features (feature attribution) can give an indication of biased prediction strategies, which we subsequently utilize for their systematic identification. We then show how we can correct models without the need for sample removal and retraining. Motivated by the reliance of our approach on feature attribution methods, we identify an issue with respect
to their robustness under model manipulation, and introduce a strategy to increase their robustness. As research prototypes are not enough to provide practitioners with an accessible approach to model debugging, we briefly introduce our fully tested and documented open source software frameworks Zennit, CoRelAy and ViRelAy, which aim to severely simplify the identification of model biases and the underlying data.
