Principled Integration of Specialized Components in Complex AI Systems
A Principled Approach to the Collaborative Integration of Specialized Components in Complex AI Systems
Funder: EPSRC
Lead partner: UCL Computer Science
Other partners: Deepmind Technologies
Lead academic: Dr Ilija Bogunovic
Project amount: £168,494
Research themes: AI and intelligent systems
Project period: 1 Nov 2023 - 31 Oct 2027
Project description: The growth of specialized AI components – models, agents, plugins, or systems, each with unique computational requirements, latency, complexity, and task specialization – necessitates a principled strategy for effective integration. Consider the challenge of integrating a dialogue system, responsible for human-like conversation, with a robotics component handling real-world interaction; this scenario exemplifies the integration tasks my research seeks to address. Leveraging a bottom-up, first principles approach, my PhD research aims to tackle key questions: 1. What optimal structure and timing for information and gradient exchange during both inference and learning stages could maximize efficiency, given individual operation frequencies, latency, and component complexity? 2. Can we develop a universal, scalable algorithm for effectively integrating diverse AI components, ensuring collaboration and alignment towards a shared objective while respecting each component's unique constraints and characteristics? 3. How can we construct a system that remains robust, reliable, safe, and adaptable in response to changes in its components or tasks? To investigate these questions, I plan to draw insights from various fields. For example, paradigms from collaborative multi-agent research could inform the alignment of AI components, hierarchical reinforcement learning could provide frameworks for managing interactions across different timescales and abstraction levels, and multimodal sequence modelling could offer insights for handling components operating at different frequencies. Through this research, I aim not only to develop more robust, adaptable, and scalable AI systems but also to contribute novel insights and methodologies to the respective fields, thereby enriching our understanding of complex, multi-component systems.
