RI: Small: Sparse Reconfigurable Artificial Neural Systems: Optimal Neuron Selection and Generalization

点击翻译按钮获取中文摘要

Machine learning and Artificial Intelligence are fueling a revolution that is making it possible to do better prediction from data, to better search for images on the internet, and even to better talk to computers using natural language. This project introduces novel machine learning methods for training artificial networks of neurons. This research also has the potential to contribute to neural science and the understanding of biological brain function. Humans display fast and flexible learning. How are our brains wired to do what they do? During normal brain development, the process of programmed cell death represents a form of "neuron selection" that helps to shape the size and configuration of different information processing centers in the brain. In effect, this wires our brain to do particular tasks. This is also thought to represent one of the most basic forms of learning. This research introduces new methods for "neuron selection" as a form of learning by machines.Current learning methods for artificial networks of neurons largely focus on adjusting signal strength between neural cells. Adjusting the strength of these signals is a slow and repetitive process. However, human learning is often spontaneous. This project looks at how artificial networks of neurons can learn by turning neurons on and off, enabling the same network to be reconfigured for multiple learning tasks. Preliminary experiments show that this can be highly effective and can result in good generalization, even when using neurons with fixed randomly generated signals. The proposed methods do not just identify "useful neurons." Instead, these methods can identify "coalitions of neurons" that work together as a team to achieve a particular goal. Neuron selection can be executed much more rapidly than learning signal strength between neurons. The proposed methods guarantee a linear time bound on learning. The methods are also guaranteed to converge to the optimal "team of neurons" relative to a given starting configuration and learning task.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.