Hardware-Software Co-Design Low-Complexity Fuzzy Algorithm for High-Dimensional Feature Space With In-Situ Learning

The exploitation of fuzzy techniques in machine learning for modelling uncertainty is highly versatile and widespread. In addition to providing a suitable fuzzy algorithm, its hardware implementation targeting proper performance in real-time applications is another challenge. In this paper, we propose a technique to deal with high-dimensional systems. Despite considering the uncertainty in the data, the proposed algorithm (HCLFA) is low-complex, thanks to the small number of calculations required. In the proposed algorithm, a D-dimensional system is transformed into a D one-dimensional system, and a Gaussian function is distributed over each training data. The main contribution is to provide a complete fuzzy system with tuning one parameter and simple operation without any complex mathematical relationships. The paper also presents an efficient memristor-crossbar hardware structure for the algorithm. The circuit is designed, and the ArC One hardware platform is used to demonstrate the results experimentally. It can learn and adapt to the application without the need for a host system and can be utilized as a basis for providing evolutionary systems. The results and comparisons indicate the significant superiority of the proposed algorithm, and achieve an accuracy of $99.67 \pm 0.25$ on the MNIST dataset.