Embedded Machine Learning: Emphasis on Hardware Accelerators and Approximate Computing for Tactile Data Processing

Machine Learning (ML) a subset of Artificial Intelligence (AI) is driving the industrial and technological revolution of the present and future. We envision a world with smart devices that are able to mimic human behavior (sense, process, and act) and perform tasks that at one time we thought could only be carried out by humans. The vision is to achieve such a level of intelligence with affordable, power-efficient, and fast hardware platforms. However, embedding machine learning algorithms in many application domains such as the internet of things (IoT), prostheses, robotics, and wearable devices is an ongoing challenge. A challenge that is controlled by the computational complexity of ML algorithms, the performance/availability of hardware platforms, and the application’s budget (power constraint, real-time operation, etc.). In this dissertation, we focus on the design and implementation of efficient ML algorithms to handle the aforementioned challenges. First, we apply Approximate Computing Techniques (ACTs) to reduce the computational complexity of ML algorithms. Then, we design custom Hardware Accelerators to improve the performance of the implementation within a specified budget. Finally, a tactile data processing application is adopted for the validation of the proposed exact and approximate embedded machine learning accelerators. The dissertation starts with the introduction of the various ML algorithms used for tactile data processing. These algorithms are assessed in terms of their computational complexity and the available hardware platforms which could be used for implementation. Afterward, a survey on the existing approximate computing techniques and hardware accelerators design methodologies is presented. Based on the findings of the survey, an approach for applying algorithmic-level ACTs on machine learning algorithms is provided. Then three novel hardware accelerators are proposed: (1) k-Nearest Neighbor (kNN) based on a selection-based sorter, (2) Tensorial Support Vector Machine (TSVM) based on Shallow Neural Networks, and (3) Hybrid Precision Binary Convolution Neural Network (BCNN). The three accelerators offer a real-time classification with monumental reductions in the hardware resources and power consumption compared to existing implementations targeting the same tactile data processing application on FPGA. Moreover, the approximate accelerators maintain a high classification accuracy with a loss of at most 5%.