Foodborne illness is a common, distressing, and even life-threatening problem for millions of people around the world. Food testing is becoming a common need for our daily lives, i.e., a sensitive and cost-effective detector can guarantee the safety and quality of food to fulfill strict food legislation and consumer demands. However, current electrochemical sensors not only require lengthy and complex sample processing, but also need expensive equipment for an adequate level of purification and enrichment. Meanwhile, low-cost sensors produce noisy signals that require computationally expensive machine-learning models for noise-robust processing. To remedy these issues, this research project is to develop an intelligent food-pathogen detection system, which is portable and affordable, yet accurate. The success of this project will enable affordable, portable and accurate food-borne pathogen detection, including food analysis, food safety, food-borne illness detection and therapy, water measurement, and many microbiome-related environmental monitoring. This project will provide an intellectual foundation for portable, disposable, intelligent foodborne pathogen detection expanding the burgeoning biosensor, machine learning and hardware accelerator design research community.
Specifically, this project aims to develop (1) a portable electrochemical sensing system detecting foodborne pathogens in contaminated foods, (2) hardware-friendly bitwise machine learning models for de-noising and recognition of real-samples and (3) an ultra-low power and portable accelerator for bit-wise machine learning inferences. The three aims are to improve the recognition performance in a collaborative fashion, where affordable sensing technology is backed up by machine-learning-based signal processing, while the hardware-algorithm co-design controls system-level efficiency.
Nanopore genome sequencing is becoming the cornerstone to enable personalized medicine, global food security and wildlife conservation. 'Base-calling' is the process of assigning bases (nucleobases) to signal peaks observed by the gene sequencer. It is the most time-consuming step during sequencing and uses deep neural networks to translate vast amounts of raw electrical signals produced by nanopore sequencers to digital DNA symbols. However, although current approaches reduce the computing overhead of a base-caller, they substantially increase uncorrectable systematic errors resulting in low accuracy for base-calling. Moreover, the low power efficiency of prior base-calling accelerators severely restrict the use of this sequencing technology in numerous real-time biomedical applications. Currently, there is no methodology to automatically explore the huge design space of a base-caller in conjunction with its hardware accelerators. To remedy these issues, this project will develop a novel algorithm & hardware co-design methodology to make nanopore base-calling more power efficient, scalable and accessible, making it possible to realize its value in socially relevant applications that demand fast genome sequencing solutions. It will also lower the barrier to nanopore sequencing development, bringing the benefits of sequencing to users who are not machine learning and hardware design experts. This project will engage undergraduate and graduate students in cutting edge interdisciplinary fields ranging from genomics to computer engineering. All artifacts and teaching materials will be broadly disseminated via open source and creative commons.
This project aims to develop and rigorously validate (1) a systematic-error-aware binarized base-caller, (2) a power-efficient spintronics accelerator for binarized base-calling, and (3) deep-reinforcement-learning-based approach to automatically explore the design space of the base-caller and its accelerator, to generate optimal frameworks for developing the next generation of base-callers. These three aims will enable even non-experts to take advantage of nanopore base-calling in numerous life science fields.