Ahmed AL-KAF

Abstract: Society is facing major challenges such as climate change and resource scarcity. The development of sustainable and energy-efficient electronic devices is crucial, especially with the increase in connected objects. 'Approximate Computing' is a promising paradigm to reduce the complexity of hardware and software components in exchange for controlled precision. This is particularly relevant for artificial intelligence (AI) applications, which can tolerate a certain degree of approximation and exhibit resilience to small perturbations. However, larger perturbations can affect the quality of results and become concerning, especially for future generations of integrated circuits.
Research Problem:
Fault tolerance and approximate computing aim to manage errors, but in different ways. Fault tolerance guarantees precise results despite defects, while approximate computing accepts results within a range of values to optimize energy efficiency. The main challenge is to find a balance between reliability and energy consumption. Reducing the precision of processed data can decrease energy consumption but also reduces the overall reliability of AI applications.
Thesis Objectives:
Design, validate, and evaluate different hardware and software implementations of neural networks using approximate computing techniques.
Evaluate their fault tolerance using fault injection techniques on FPGA boards, covering both learning and inference phases.
Identify the best synergies between approximate computing and fault tolerance, and propose optimal design methodologies in terms of reliability, energy consumption, and quality of results.