Energy-efficient embedded machine learning algorithms for smart sensing systems

Embedded autonomous electronic systems are required in numerous application domains such as Internet of Things (IoT), wearable devices, and biomedical systems. Embedded electronic systems usually host sensors, and each sensor hosts multiple input channels (e.g., tactile, vision), tightly coupled to the electronic computing unit (ECU). The ECU extracts information by often employing sophisticated methods, e.g., Machine Learning. However, embedding Machine Learning algorithms poses essential challenges in terms of hardware resources and energy consumption because of: 1) the high amount of data to be processed; 2) computationally demanding methods. Leveraging on the trade-off between quality requirements versus computational complexity and time latency could reduce the system complexity without affecting the performance. The objectives of the thesis are to develop: 1) energy-efficient arithmetic circuits outperforming state of the art solutions for embedded machine learning algorithms, 2) an energy-efficient embedded electronic system for the “electronic-skin” (e-skin) application. As such, this thesis exploits two main approaches: Approximate Computing: In recent years, the approximate computing paradigm became a significant major field of research since it is able to enhance the energy efficiency and performance of digital systems. “Approximate Computing”(AC) turned out to be a practical approach to trade accuracy for better power, latency, and size . AC targets error-resilient applications and offers promising benefits by conserving some resources. Usually, approximate results are acceptable for many applications, e.g., tactile data processing,image processing , and data mining ; thus, it is highly recommended to take advantage of energy reduction with minimal variation in performance . In our work, we developed two approximate multipliers: 1) the first one is called “META” multiplier and is based on the Error Tolerant Adder (ETA), 2) the second one is called “Approximate Baugh-Wooley(BW)” multiplier where the approximations are implemented in the generation of the partial products. We showed that the proposed approximate arithmetic circuits could achieve a relevant reduction in power consumption and time delay around 80.4% and 24%, respectively, with respect to the exact BW multiplier. Next, to prove the feasibility of AC in real world applications, we explored the approximate multipliers on a case study as the e-skin application. The e-skin application is defined as multiple sensing components, including 1) structural materials, 2) signal processing, 3) data acquisition, and 4) data processing. Particularly, processing the originated data from the e-skin into low or high-level information is the main problem to be addressed by the embedded electronic system. Many studies have shown that Machine Learning is a promising approach in processing tactile data when classifying input touch modalities. In our work, we proposed a methodology for evaluating the behavior of the system when introducing approximate arithmetic circuits in the main stages (i.e., signal and data processing stages) of the system. Based on the proposed methodology, we first implemented the approximate multipliers on the low-pass Finite Impulse Response (FIR) filter in the signal processing stage of the application. We noticed that the FIR filter based on (Approx-BW) outperforms state of the art solutions, while respecting the tradeoff between accuracy and power consumption, with an SNR degradation of 1.39dB. Second, we implemented approximate adders and multipliers respectively into the Coordinate Rotational Digital Computer (CORDIC) and the Singular Value Decomposition (SVD) circuits; since CORDIC and SVD take a significant part of the computationally expensive Machine Learning algorithms employed in tactile data processing. We showed benefits of up to 21% and 19% in power reduction at the cost of less than 5% accuracy loss for CORDIC and SVD circuits when scaling the number of approximated bits. 2) Parallel Computing Platforms (PCP): Exploiting parallel architectures for near-threshold computing based on multi-core clusters is a promising approach to improve the performance of smart sensing systems. In our work, we exploited a novel computing platform embedding a Parallel Ultra Low Power processor (PULP), called “Mr. Wolf,” for the implementation of Machine Learning (ML) algorithms for touch modalities classification. First, we tested the ML algorithms at the software level; for RGB images as a case study and tactile dataset, we achieved accuracy respectively equal to 97% and 83.5%. After validating the effectiveness of the ML algorithm at the software level, we performed the on-board classification of two touch modalities, demonstrating the promising use of Mr. Wolf for smart sensing systems. Moreover, we proposed a memory management strategy for storing the needed amount of trained tensors (i.e., 50 trained tensors for each class) in the on-chip memory. We evaluated the execution cycles for Mr. Wolf using a single core, 2 cores, and 3 cores, taking advantage of the benefits of the parallelization. We presented a comparison with the popular low power ARM Cortex-M4F microcontroller employed, usually for battery-operated devices. We showed that the ML algorithm on the proposed platform runs 3.7 times faster than ARM Cortex M4F (STM32F40), consuming only 28 mW. The proposed platform achieves 15× better energy efficiency than the classification done on the STM32F40, consuming 81mJ per classification and 150 pJ per operation.