Big Data/Machine Learning/AI
A Comparison of Deep Neural Network Compression for Citizen-Driven Tick Surveillance Yichao Liu* Yichao Liu Emmanuel Dufourq Joacim Rockloev
Vector-borne diseases, like Crimean-Congo hemorrhagic fever (CCHF) from ticks, pose global health risks. Citizen science initiatives use web and mobile apps to enable public reporting of vector sightings, enhancing data collection for monitoring populations and disease spread. Machine learning can automate species identification, improving surveillance and reducing manual effort. However, deploying these models on mobile devices is challenging due to their over-parameterized nature, requiring significant storage, computation, and energy—constraints for battery-limited devices. Effective model compression is essential for practical deployment in citizen science applications.
In this study, mainstream model compression techniques like pruning and quantization were evaluated to reduce the size of object detection models, as object detection acts as the first step in identification and has a strong impact on overall performance. Pruning removes non-essential parameters to lower computational demands and improve speed. Two types of pruning were tested: structured, which removes entire model sections, and unstructured, which selectively removes individual elements.
Quantization reduces model size by using lower-bit formats (e.g., 8-bit, 4-bit) instead of 32-bit floating-point numbers. Dynamic quantization adjusts precision during processing, while static quantization converts parameters to lower precision before deployment, ensuring consistent size and speed improvements.
The impact of compression techniques was evaluated on mobile devices by comparing the runtime performance of compressed and original models under identical conditions. Combining structured and unstructured pruning reduced model size by up to 60% with minimal performance loss, addressing limitations of each method alone. Quantization also proved effective, significantly reducing size and runtime with negligible performance impact.
Effective model compression techniques enable real-time, automated identification of disease vectors on mobile devices. This study’s findings could enhance public participation in vector surveillance, aiding in the control of vector-borne diseases.
