Optimal Parameters Determination for Extreme Learning Machine in the Human Activity Recognition

Extreme Learning Machine is a single-hidden-layer feed-forward neural network that has the advantage of high learning speed and ease of implementation. The efficiency of an extreme learning machine in human activity recognition largely depends on three parameters, such as the weight matrix, the hidden layer neurons number, and the activation functions. This study is aimed at building an artificial neural network to solve the problem of human activity recognition to analyze the influence on the accuracy of parameters neural networks such as input weights, activation functions, and the neurons number in the hidden layer. For the experiment, an open dataset was used, which includes information about seven physical activity types. We compared the accuracy when training a neural network with an extreme learning machine and an extreme learning machine with the weight coefficient values calculated using the particle swarm optimization method. Also, the influence on the accuracy of such activation functions as a hyperbolic tangent, rectified linear unit, sigmoid, sinusoidal, and binary step function for different hidden layer neurons numbers was evaluated.