Random Kernel Perceptron on ATTiny2313 Microcontroller

Kernel Perceptron is very simple and efficient online classification algorithm. However, it requires increasingly large computational resources with data stream size and is not applicable on large-scale problems or on resource-limited computational devices. In this paper we describe implementation of Kernel Perceptron on ATTiny2313 microcontroller, one of the most primitive computational devices with only 128B of data memory and 2kB of program memory. ATTyny2313 is a representative of devices that are popular in embedded systems and sensor networks due to their low cost and low power consumption. Implementation on microcontrollers is possible thanks to two properties of Kernel Perceptrons: (1) availability of budgeted Kernel Perceptron algorithms that bound the model size, and (2) relatively simple calculations required to perform online learning and provide predictions. Since ATTiny2313 is the fixedpoint controller that supports floating-point operations through software which introduces significant computational overhead, we considered implementation of basic Kernel Perceptron operations through fixed-point arithmetic. In this paper, we present a new approach to approximate one of the most used kernel functions, the RBF kernel, on fixed-point microcontrollers. We conducted simulations of the resulting budgeted Kernel Perceptron on several datasets and the results show that accurate Kernel Perceptrons can be trained using ATTiny2313. The success of our implementation opens the doors for implementing powerful online learning algorithms on the most resource-constrained computational devices.