Sustainable AI Solutions to Road Safety and Incremental Learning Problems

The digitization of services and processes across industries has led to a sub- stantial increase in global electricity demand and a significant driver to this trend is artificial intelligence (AI) technologies. The energy-intensive train- ing and deployment of machine learning models, which are ubiquitous in real-world applications, present significant challenges to sustainability. This dissertation aims to contribute to sustainable and safety-enhancing AI in two specific areas: road safety and continual learning. The first part of this dissertation explores the integration of artificial intelligence into Advanced Driver Assistance Systems (ADAS) to enhance road safety, with a focus on Relevant Trac Light Recognition. Conducted in collaboration with Verizon Connect, this research explore the challenge of identifying trac lights relevant to the ego-vehicle in images captured by vehicle-mounted cameras. The second part focuses on Continual Learning (CL), a machine learning paradigm that enables models to incrementally learn from new data while avoiding catastrophic forgetting. This dissertation introduces novel methods for exemplar-free class-incremental learning, namely Elastic Feature Consoli- dation (EFC) and its enhanced version, EFC++, which mitigate feature drift through innovative regularization techniques. This second part concludes with an in-depth analysis of the energy consumption of selected pre-trained class-incremental learning strategies. The findings help identify processes used to mitigate forgetting that have minimal impact on energy consump- tion during both training and inference phases, thereby enhancing scalability and promoting sustainability.