Integrating Cognitive Architecture and Neural Generative Language Models
Abstract
Possibly the most significant challenge in supporting high-quality, real-time interaction between Artificial Intelligence (AI) agents and humans is communication. Although significant progress has been achieved in many aspects of natural language, advances have mostly been limited to controlled settings. Existing systems do not extract internal representations of the underlying semantics and pragmatics from language grounded in an environment shared with humans, nor can they productively reason with those representations when interacting with humans. Neural generative language models offer tantalizing promise for improving natural language understanding for AI systems. They have demonstrated an almost eerie ability to create coherent text that suggests a deep understanding of a subject, in spite of the fact that they have no internal models of the world. These neural models have not yet been incorporated within more complete cognitive agents to support interaction with humans. Further, there are no established methods for extracting structured knowledge from these models in ways that assure the providence and reliability of the knowledge. In contrast, research on cognitive architectures has made continual progress in creating frameworks for creating flexible and robust multi-task agents that encodestructured knowledge and integrate a wide range of complex cognitive capabilities, including human-machine interaction. We propose to pursue a novel, targeted research integration of these AI technologies, emphasizing both what cognitive architectures do best (support end-to-end integration of interaction, reasoning, language processing, learning, etc. using structured, curated knowledge) andwhat deep neural generative language architectures do best (provide associational retrieval from massive stores of latent unstructured, possibly unreliable knowledge). We will develop methodologies so that as tens to hundreds of millions of dollars are invested in new neural architectures, we can exploit them through integration with cognitive architectures.We will use the Soar cognitive architecture and extend its capabilities by integrating one or more neural generative language models. We will explore how the latent knowledge and completion capabilities of generative language models can improve language processing and interactive task learning more.velopment of general techniques and theories that can be shared with other researchers in cognitive systems. Thus, a potential outcome is a novel scientific foundation with the potential to speed and to ease the exploitation of neural language models for the larger community of cognitive systems researchers. The new capabilities offered by the envisioned research offer the potential for sficant improvements in the capability of deployed intelligent systems for the DoD, especially those that interact, collaborate, andcooperate with humans. Todays solutions lack a combination of scalability, reliability, and flexibility. However, this combinationis critically needed for most DoD problems. Reliability and trustworthiness are essential for confident and ready use of operational systems. The research approach we propose will support autonomous agents that can be flexible to changing tasks in support of new or different missions and to seek out, evaluate, and consolidate the new knowledge required for these missions. Potential applications (among many) include support for planning, adaptive monitoring of the environment (helping build and sustain common operating picture), and multi-domain operations (MDO; e.g., rapid adaptation to adversary tactics based on captured reports).
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 05, 2021
- Source ID
- N000142112369
Entities
People
- John E. Laird
Organizations
- Office of Naval Research
- United States Navy