Description

This fellowship opportunity at the National Institute of Standards and Technology (NIST) focuses on advancing machine learning by developing training, optimization, and benchmarking methods for hardware neural networks. Traditional machine learning algorithms, such as deep learning, are computationally intensive and power-hungry when run on conventional digital computers. To overcome these limitations, emerging AI hardware platforms utilize diverse materials and technologies—including photonic, memristive, magnetic, and superconducting devices—often employing analog or mixed-signal processing to achieve higher speeds and energy efficiency.

Despite these advances, a major challenge remains: the incompatibility of traditional training algorithms like backpropagation with these novel hardware architectures. This project aims to create and demonstrate a general training technique that can be natively implemented across this variety of hardware neural networks. Research areas include:

• Developing physical models for different hardware platforms.
• Extending training techniques to spiking neural hardware such as Intel's Loihi and commercial FPGAs.
• Implementing continual learning and few-shot learning benchmark tasks to evaluate performance.

This research is expected to contribute to more efficient, scalable AI hardware systems capable of overcoming current limitations in power consumption and algorithmic compatibility.