Hardware-Software co-design Methodologies for Edge AI Optimization

ML-based edge devices may face memory and computational errors that affect applications' reliability and performance. These errors can be the result of particular working conditions (e.g., radiation areas in physical experiments or avionics) or could be the consequences of energy/power optimization approaches. In this context, memories are particularly affected, because their large contribution in the total energy cost made the research community focus on them to find energy-aware solutions. On the same line, approximate computing reduces energy consumption at the cost of inexact results. The exploration of robust designs to mitigate the impact of these errors in ML-based embedded system is the main objective of this research. Indeed, the trade-off between computational accuracy and energy saving constitutes the core exploration in this thesis proposal. Although existing models like convolutional neural networks are known to be quite resilient to noise, specific design can improve their resiliency.In parallel to the error robustness area of research, energy saving approaches will be investigated as a complementary task. Here, the usage of hardware accelerators constitutes a key point. Additionally, an exploration of the effect of errors in the target accelerators completes the range of possible case studies in the field of error resiliency.To conduct such a research, suitable memory and computational models are necessary to simulate the error impacts at the application level. Finally, the proposed solutions should be evaluated in state-of-the-art (SoA) applications to estimate their benefits in real life scenarios. The conducted research will investigate biomedical applications, with a clear focus on wearable devices for patient monitoring.