TrinityAI: towards trustworthy, resilient, and interpretable AI for high-assurance applications
Susmit Jha · 2024
In this talk, we will describe TrinityAI - a neuro-symbolic artificial intelligence (AI) architecture being developed for detecting complex events across varied environments. Complex events are scenarios composed of multiple atomic events that occur over different times and locations, involving various actors. TrinityAI aims to enhance monitoring and situational awareness in critical applications such as healthcare, surveillance for public security, and traffic monitoring. The desiderata of AI needed to detect such complex events include the following: first, we need multimodal perception in very high dimension; second, we need to reason to compose atomic events and infer complex event; and finally, we need the capability to ingest background knowledge from human experts, and provide symbolic justification of AI decisions back to the human users. TrinityAI satisfies these requirements and its architecture is motivated by predictive processing - a theory of mind which posits that human cognition is not bottom-up wherein we attach meanings to our perceptions in isolation from each other but instead we build models or narratives of the world and use that to interpret our sensed perceptions. This model-based perception is key to the robustness and context-sensitivity of our sensing and decision-making. TrinityAI has a three-layered architecture composing deep neural networks, foundation models, and logic programming with each layer designated to handle specific tasks ranging from raw data perception to high-level reasoning and inference.