HAC-M-DNN: Hardware Aware Compression of Sustainable Multimodal Deep Neural Networks for Efficient TinyML Deployment

The increasing complexity and energy consumption of advanced artificial intelligence algorithms pose a significant challenge to environmental sustainability. This paper addresses this issue by introducing HAC-M-DNN, a system designed for energy-efficient tinyML deployment. HAC-M-DNN effectively manages multimodal data, develops carbon footprint-efficient multimodal deep neural networks, and employs hardware-aware model compression techniques to optimize memory utilization and power efficiency. We validated HAC-M-DNN through two tinyML application case studies: semantic segmentation of crop and weed species using multimodal RGB-D imaging and scene understanding from multimodal image and audio data, achieving accuracies of 70% and 95%, respectively. Our hardware-aware results demonstrate that HAC-M-DNN generates highly accurate, compact models deployable on resource-constrained devices, achieving up to 85× and 177× reduction in memory usage respectively for two case-studies. These models also exhibit low latency and exceptional power efficiency. The outcomes of this study underscore the potential of HAC-M-DNN to drive both technological advancement and environmental responsibility in the field of tinyML, offering a sustainable alternative to traditional, resource-intensive AI models.