Research

This paper presents a novel approach to optimize deep learning models for deployment in environments with limited computational resources. We introduce a compression technique that reduces model size by up to 80% while maintaining 95% of the original accuracy.

We propose a framework for building scalable distributed systems capable of processing large volumes of data in real-time. Our approach leverages a novel partitioning strategy that reduces latency by 40% compared to state-of-the-art methods.

Developing advanced multimodal learning techniques that combine imaging data, electronic health records, and genomic information for improved medical diagnosis and treatment planning.

This paper addresses the challenge of training machine learning models on sensitive healthcare data while preserving patient privacy. We introduce a federated learning approach with differential privacy guarantees that enables collaborative model training without sharing raw patient data.

We present a framework for making complex AI models more interpretable and explainable, particularly for applications in critical decision-making domains. Our approach combines post-hoc explanation methods with inherently interpretable model architectures to provide comprehensive explanations for model predictions.

The Locate API is designed to optimize spatial site selection by identifying the best locations for sensor deployment within a given polygon. This research aims to evaluate the efficiency and accuracy of the Locate API in selecting sites that meet specific spatial constraints, such as mandatory locations, minimum distance requirements, and land-use considerations. The study will also explore the potential improvements and applications of the API in real-world scenarios.

Urban air quality monitoring is critical for public health, particularly in rapidly growing cities like Kampala, Uganda, where pollution levels rise alongside urbanization. The AirQo network, deploying over 60 low-cost PM2.5 sensors, provides vital air quality data but faces challenges in balancing spatial coverage, temporal reliability, and operational costs. This study proposes a multi-method optimization framework to enhance the network’s efficiency. First, spatial distribution analysis employs spatial autocorrelation (Moran’s I, Getis-Ord Gi*) and Voronoi diagrams to evaluate coverage gaps and redundancies. The second, temporal analysis investigates diurnal/seasonal PM2.5 trends and device redundancy via cross-correlation and downtime assessments. Third, machine learning and clustering (K-means, feature importance, simulated annealing) optimize sensor placement to minimize redundancy while preserving data accuracy. Finally, cost-benefit analysis quantifies financial savings and data quality trade-offs, incorporating stakeholder feedback to ensure policy relevance. This research aims to deliver actionable strategies for cost-effective, high-fidelity air quality monitoring by integrating spatial, temporal, computational, and economic perspectives. The findings will guide policymakers in Kampala and similar cities to prioritize sensor deployment, reduce operational costs, and strengthen evidence-based environmental management. This holistic approach underscores the value of interdisciplinary methodologies in addressing urban sustainability challenges.