BrIAS Seminars: Mehrdad Asad скачать в хорошем качестве

BrIAS Seminars: Mehrdad Asad

27.03.2025 BrIAS Junior Fellow Dr. Mehrdad Asad Augmenting Machine Learning Pipeline with Explainability Abstract: Over the years, with the advancements in computing infrastructure, tremendous amounts of data generated by software systems have been fueled by Artificial Intelligence (AI), in particular, Machine Learning (ML), to generate actionable insights. These advancements pose new concerns such as data quality, data bias, etc. compared to classical software architecture. Additionally, with the complexity arising from applying sophisticated AI and ML algorithms, a key limitation for the adoption of AI on a scale is its inherent black-box characteristics. In this talk, we present a set of new architectural requirements for such ML-based system software architecture and present proof of concept system architecture that augments AI components with explainability methods to monitor the inference capabilities. We will walk through use cases in adversarial AI and possible challenges with different XAI methods posing the needs for lightweight and efficient XAI techniques for large-scale deployment.