UKRI/BBSRC-NSF/BIO: Interpretable and Noise-Robust Machine Learning for Neurophysiology

University Of California-Irvine, Irvine CA

Investigators

Abstract

Recent advancements in machine learning have revolutionized various fields, including the biomedical sciences. However, the adoption of modern machine learning techniques in neuroscience remains limited. Given that neuroscience involves identifying nonlinear systems by utilizing experimental observations to characterize them, machine learning has the potential to be transformative in this domain. Over the past decade, significant progress has been made in developing powerful experimental techniques that enable the observation of neural signals at larger scales and higher resolutions than ever before. Unfortunately, the conceptual progress in the field has been slow due to the lack of corresponding advancements in data analysis approaches. Many studies still rely on classical tools that overlook the richness and complexity of neural signals. The objective of this project is to create a toolkit that empowers systems neuroscientists to construct models connecting sensation, perception, and cognition. The proposed framework enables neuroscientists to fit flexible models capable of performing essential perceptual and cognitive tasks directly from neural recordings. The significance of this research lies in its potential to enhance our understanding of the brain's complex functions, which can ultimately lead to the development of advanced therapeutic methods and diagnostic tools for neurological disorders. Moreover, it will provide valuable insights into human cognition, potentially enhancing artificial intelligence and machine learning applications. This project also aims to establish an integrated educational and outreach plan, including interdisciplinary courses and programs accessible to undergraduate and graduate students from computer science, cognitive science, and the school of medicine. This research project focuses on the development of a biologically plausible and interpretable modeling framework that employs neuro-symbolic representations to offer a hierarchical explanation of perception and cognition. The proposed framework bridges the gap between machine learning and neuroscience, opening up new avenues for understanding and interpreting brain function. It comprises two main stages: (1) an encoding stage that models data transformation through spiking neurons, emulating the anatomy and physiology of early sensory pathways, and (2) a cognitive stage that establishes neuro-symbolic models using neural representations and algorithms that simulate higher-level brain dynamics. The cognitive stage will adhere to biologically plausible computations, facilitating the interpretation of model phenomena at a mechanistic level. To validate the framework, neural signals will be recorded across various scales and resolutions (from single units to EEG) within the context of hearing. Furthermore, behavioral experiments will be conducted to evaluate the model's ability to replicate human behavior in common perceptual and cognitive tasks. All developed tools will be released as open-source libraries, serving as valuable resources for the neuroscience community, including non-experts in modeling or programming. This broad accessibility not only facilitates the proliferation of knowledge but also encourages the development of innovative solutions in the field, further enhancing its societal impact. 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.