Neuromorphic Computing for Low-Power AI in Ghana - Research

Abstract

The high energy consumption of conventional AI models is a major barrier to their widespread deployment in energy-constrained environments, such as off-grid rural areas in Ghana. Neuromorphic computing, which draws inspiration from the brain's architecture to build highly efficient hardware, offers a radical solution. This Masters in Computer Science proposal focuses on the development and application of a neuromorphic system for a critical edge computing task in Ghana: the real-time, on-device diagnosis of crop disease from leaf images using a spiking neural network (SNN). The research aims to demonstrate an order-of-magnitude reduction in power consumption compared to traditional AI on a GPU or TPU, enabling continuous, solar-powered monitoring in the field.

Key Research Questions for Neuromorphic AI in Ghana

Event-Based Sensing and Processing: How can we leverage event-based cameras, which only transmit data when a pixel's brightness changes, to create an ultra-low-power pipeline for monitoring plant health? This contrasts with traditional frame-based cameras that consume constant power.
SNN Training for Real-World Data: What is the most effective method to train a Spiking Neural Network for a practical image classification task? This research will compare surrogate gradient descent methods with unsupervised learning approaches like Spike-Timing-Dependent Plasticity (STDP).
Hardware-Algorithm Co-design: How can an SNN algorithm be co-designed with the constraints of a specific neuromorphic chip (e.g., Intel's Loihi 2) to maximize accuracy and minimize energy-per-inference?
Field Deployment and Robustness: How robust is a neuromorphic system to real-world conditions in a Ghanaian farm, including variable lighting, camera movement, and different stages of crop growth?

Proposed Masters in Computer Science Research: A Neuromorphic Cassava Disease Detector

This thesis will focus on building a complete, low-power system for identifying Cassava Mosaic Disease and Cassava Brown Streak Disease, two major threats to food security in Ghana.

Dataset Collection: An event-based dataset will be created by recording videos of healthy and diseased cassava plants in Ghana using a specialized event-based camera. This will be one of the first datasets of its kind for agricultural applications.
SNN Model Development: A deep SNN will be designed and trained to classify the event-based data. The network will be trained to be robust to changes in scale and orientation, a common challenge in field imaging.
Deployment on Neuromorphic Hardware: The trained SNN will be deployed on a research-access neuromorphic platform (such as Intel's Neuromorphic Research Community). The system's power consumption and inference speed will be meticulously benchmarked against a conventional CNN running on a mobile GPU (e.g., a Jetson Nano).
Prototype Field Device: A prototype device will be assembled, integrating the event-based camera and neuromorphic hardware into a small, solar-powered unit that could be placed in a field for continuous, autonomous monitoring.

Impact for Ghana and Africa

This research will pioneer the application of neuromorphic computing to solve a critical problem in African agriculture. A successful project would demonstrate a new paradigm for AI at the edge: intelligent, autonomous sensors that can operate for months or years on a small solar panel and battery. This "deploy-and-forget" capability could be revolutionary for a wide range of applications in Ghana and across Africa, from wildlife poaching detection and environmental monitoring to remote patient health tracking. By building expertise in this next-generation AI hardware, Ghana can position itself to leapfrog traditional, energy-intensive AI infrastructure and become a leader in sustainable, brain-inspired computing.