Using biologically hierarchical modular architecture for explainable, tunable, generalizable, spatial AI
Abstract
Explainable artificial intelligence (XAI) is an area of ongoing research for a variety of machine learning applications that aims to describe decision making employed by many artificial neural networks (ANNs) in an intuitive manner for humans. Another seemingly unsolved topic generating research interest is AI conceptual understanding, (such as in large language models) which falls within a more general problem of AI model generalizability and integrability. How might computational problems of that difficulty be solved, and more importantly, which approach will result in genuine advancement in the fields of neuroscience and/or artificial intelligence? An oftenused strategy in machine learning is to use knowledge about biological brains to inspire machine learning model design. That strategy can continue to be useful and novel because both neuroscience and artificial intelligence are rapidly growing fields of research; the same unification strategy between both fields can yield significantly different models as neuroscience knowledge and machine learning capabilities both grow and develop. In this paper, we focus on the usefulness of using current neuroscience knowledge to design biologically hierarchical and modular architecture (BHMA) models of the brain that can solve spatial learning tasks. We discuss the potential implications of biologically constrained hierarchical neural networks for the future of human-computer interaction, computational neuroscience, and machine learning given their potential for generating more explainable models with verifiable aspects of conceptual understanding.
- Publication:
-
Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series
- Pub Date:
- June 2023
- DOI:
- 10.1117/12.2679920
- Bibcode:
- 2023SPIE12542E..0WJ