CRII: RI: Neuromodulated Deep Vision: Continual Learning and Goal-Oriented Adaptation

点击翻译按钮获取中文摘要

Deep neural networks are revolutionizing many fields and industries. These systems can learn to solve problems ranging from parsing natural language to discovering new drugs. However, current networks can only solve individual, static problems; they cannot learn gradually over time or adapt their computations in real time. This project seeks to bridge the gap between human and machine intelligence by modeling a chemical process in the brain that help humans learn and adjust their behavior based on context. Specifically, the chemical acetylcholine (ACh) has been shown to modulate responses to visual stimuli by selectively raising or lowering the activity level of individual neurons. This chemical is also crucial for visual learning in infants. An insufficient concentration of ACh in the brain leads to poor visual attention and has also been linked to memory deficits and Alzheimer?s disease. Thus, adding ACh-like mechanisms to deep neural networks can potentially allow them to learn over time and adapt their behavior to rapidly changing conditions--a growing need in domains ranging from forest conservation to cybersecurity.In more detail, deep convolutional neural networks (CNNs) are the state-of-the-art approach for a wide variety of image processing tasks, such as object recognition and semantic segmentation. This project will investigate how to add ACh-like regulation to CNNs by modeling the spatial and temporal dynamics of this chemical process. Specifically, the CNN will be connected to a set of spatially and temporally heterogeneous sources of modulation. Each neuron in the CNN will have different receptors that determine how that particular neuron responds to different ambient levels of ACh. The system will use two controllers: one to determine the current level of ACh across different parts of the network and another to set the correct context (i.e., to model top-down feedback from higher cortical areas). The first controller will allow the system to quickly adjust its behavior, while the second will enable long-term learning. This project constitutes an initial step in developing intelligent systems that, like biological organisms, can autonomously cope with changing, dynamic environments. In addition to its technical contributions, this project will allow students at Georgia State University, one of the most diverse institutions in the country, to foster interdisciplinary connections between artificial intelligence, neuroscience, and related STEM disciplines.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.