Combining Acceleration and Approximation in Dependable Edge AI: Optimization Methodology and Tools Applied to a Case Study

Edge AI hardware/software implementations are increasingly required in many application areas. In the most demanding cases, applicative features must be reached with multi-criteria optimizations targeting computation speed and efficiency, power/energy reduction and dependability qualification. Approximate computing (AxC) is one common approach to achieve some of the required optimizations in Artificial Neural Networks. We discuss first in this paper an approach combining AxC and hardware acceleration, so that the best benefits can be obtained from both. A demonstrator is implemented on a SoC FPGA with soft CPUs associated or not to an accelerator. Then the focus is put on dependability aspects. A fault-injection framework implemented on the same platform allows analyzing the robustness of the demonstrator. Some global trends are shown and discussed, including surprising outcomes about the most critical bit-flips observed during the experiments. Challenges faced by designers are discussed. All results come from a case study based on an integer version of LeNet performing digit recognition.