Filip and his team investigate how to transform the development of human-centric AI with common sense to ultimately support social good applications. They conduct fundamental research on commonsense reasoning, multimodality, and analogy, and investigate the societal impact of commonsense AI on online social safety.
Their approach integrates four key pillars: theory, background knowledge, explainable decision-making, and theory-grounded benchmarking. They draw on theories from cognitive science, communication science, philosophy of AI, and linguistics.
Operationalizing the theories emphasizes the need to represent broad, multi-perspective background knowledge. Explainable decision-making is driven by methods that learn and reason: these leverage neural components for perception (typically foundational language and vision models) and grounding in theories and background knowledge. Example methods include entailment trees, prototype networks, and case-based reasoning techniques.
Ultimately, to enhance the representativeness of evaluations and enable fair assessment of state-of-the-art techniques, they develop theory-grounded evaluation frameworks and benchmarks. A key ‘secret sauce’ that makes this ambitious enterprise achievable is collaboration with partners diverse in terms of expertise and interests.
