A Performance Study of Deep Neural Network Representations of Interpretable ML on Edge Devices with AI Accelerators.

Area of Science:

• Edge computing
• Machine learning
• Deep learning

Background:

• Increasing adoption of machine learning (ML) and deep learning (DL) drives demand for edge deployment in sensor applications.
• Current ML implementations on edge devices are often inefficient, complex, and lack interpretability.

Purpose of the Study:

• To develop a novel, application-oriented approach for representing interpretable ML inference as deep neural networks (DNNs).
• To enhance latency and energy efficiency of ML algorithms on edge hardware for smart sensor applications.
• To enable generic artificial intelligence (AI) accelerators for interpretable ML algorithms.

Main Methods:

• Integrated interpretable deep neural network representation (IDNNRep) into an open-source ML toolbox.
• Applied the approach to regression and classification tasks in predictive maintenance (PM).
• Validated inference on edge hardware (NPU and TPU) and compared performance across quantization levels against Python and C++ implementations.

Main Results:

• Reduced inference time by up to 80% and mean energy consumption by up to 76% compared to C++ implementations, with minimal accuracy loss (0.4%).
• Achieved further performance gains using generic AI accelerators, resulting in a 94% reduction in both inference time and mean energy consumption.
• Demonstrated improved efficiency and interpretability for ML algorithms on edge devices.

Conclusions:

• The novel IDNNRep approach enables efficient, interpretable ML inference on edge hardware using generic AI accelerators.
• This method significantly improves performance for smart sensor applications like predictive maintenance and condition monitoring.
• The findings pave the way for wider adoption of efficient and interpretable ML in resource-constrained environments.