Collaborative Research: Neural computational rules of robust and generalizable learning

Michigan State University, East Lansing MI

Investigators

Abstract

Living organisms can learn from a few examples and apply that knowledge to new situations. For instance, humans can learn about trees from a few instances and recognize trees of different shapes and sizes, as well as can identify trees during different seasons, times of day, and scenes. This ability to learn fast and generalize to broader situations is unique to biological systems. In contrast, current artificial intelligence models can only attain close to human-level performance if they are trained on all possible scenarios they would encounter in each use case. This level of training is impractical, inefficient, and unrealistic. This project aims to learn directly from biological brains to identify new learning rules and apply them to artificial intelligence. The investigators train live insects, record neuronal signals from their brains, and employ computational models to identify key learning rules for fast, reliable and efficient learning. These biological principles inform the development of novel powerful algorithms for AI systems that can learn quickly, transfer knowledge to new tasks, and be robust and efficient. AI is essential in modern society, from healthcare to national security. This project leads to fundamental basic science discoveries as well as significant societal impacts. Additionally, the researchers establish summer workshops for high school, undergraduate, and graduate students, providing hands-on research experience. Both investigators are committed to training underrepresented minority students through this project. Associative learning is a crucial adaptive mechanism that influences behavioral outcomes in human and animal life. Biological systems exhibit the ability to generalize learned stimuli to diverse contexts, even from a small number of examples. However, the fundamental neural computations underlying fast, robust, and generalizable learning in biological systems are not fully understood. There exists a knowledge gap regarding the contribution of upstream neural circuits to system-level learning and the extension of biophysical learning rules for the development of new artificial intelligence (AI) algorithms.The invertebrate olfactory system shares organizational and functional similarities with the human olfactory system, making it an ideal model for investigating generalizable associative learning rules. In this study, the investigators uncover the neural computational rules governing generalizable associative learning in the central circuitry of the locust olfactory pathway and their connection to behavioral outcomes. Specific objectives are as follows: (a) Determine changes in neural responses and neural correlates of generalizable learning in the locust antennal lobe induced by associative learning; (b) Identify potential learning rules for generalization using a novel computational machine learning approach; (c) Validate the derived learning rules from the computer model through behavioral experiments. Findings from these objectives enhance understanding of the fundamental principles underlying generalizable associative learning and identify canonical AI algorithms for robust and generalizable learning. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.