Random forest-LNS architecture and vision

Efficient hardware implementations of machine-learning techniques yield a variety of advantages over software solutions: increased processing speed, and reliability as well as reduced cost and complexity. In this paper RF technique is modified so that classification is performed by LNS arithmetic. The model is applied for generic object recognition task, it shows that at low precision the RF-LNS hardware has significant area savings compared to the fixed-point alternative. With these characteristics, RF-LNS may be a good way for designing a real-time low power object recognition systems. Our future goals include further exploring precision requirements for hardware RF-LNS, noise analysis to determine the robustness of the hardware classifier and expanding LNS hardware architectures to other machine learning algorithms. Dataset Bikes Tree Units 16-bit LNS 10-bit FX 20-bit FX 3 315 219 576 4 498 407 713 5 611 622 878 6 823 835 1103 7 1010 974 1345 Cars 3 277 283 603 4 397 476 783 5 536 694 866 6 784 943 1002 7 989 1287 1311 Persons 3 336 318 409 4 534 535 657 5 765 689 845 6 878 926 1127 7 1123 1158 1287