SVM-Like Algorithms and Architectures for Embedded Computational Intelligence

This paper considers implementation of computational intelligence paradigms on resource-constrained platforms, an issue of the day for the age of invisible computing and smart sensors. This necessitates the development of ``light'' yet efficient hardware-friendly algorithms relying on a scanty power supplies and able to operate in stand-alone manner. The scope of the presented work is twofold. Firstly, we propose a new SVM-like approximated algorithm suitable for embedded systems due to good robustness and sparsity properties. Support vectors are considered as parameters of the outer optimization problem, whereas the inner one is solved using the primal representation, which reduces computational complexity and memory usage. Along with classical Gaussian kernel, a recently proposed hardware-friendly kernel, whose calculation requires only shift and add operations, is considered. Experimental results on several well-known data sets demonstrate the validity of the proposed approach, which in many cases outperforms the original RSVM using the same number of vectors. Secondly, we implement such kind of algorithm on a resource-constrained device such as a simple 8-bit microcontroller. The case-study considered further in this work is the design of a node of a wireless video-sensor network performing people detection, and a simple resource-constrained FPSLIC-based platform from Atmel is considered as a target device.