Sensor calibration is vital to have valid measurements of physical activities. In this paper, we deal with adjusting the signal from a wearable force sensor against a reference scale. By using a few samples and data augmentation, we trained a neural-based regression model to correct the wearable output. For this task, we tested the novel Auto-Rotating Perceptrons (ARP). We found that a neural ARP model with sigmoid activations can outperform an identical neural network based on classic perceptrons with sigmoid and even ReLU activation.
When changing classic perceptrons to ARP, the test loss of the sigmoid networks was reduced by a factor of 15 at the cost of increasing the execution time by ∼12% (see bar graph below).
Erratum: In the oral presentation video, when I say “without learning the inference structure of the perceptron”, I mean “without altering the inference structure learned by the perceptron”.
This paper proposed an improved design of the perceptron unit to mitigate the vanishing gradient problem at deep neural networks. The results show that models with Auto-Rotating Perceptrons (ARP) can achieve better learning performance than equivalent networks with classic perceptrons.
The modification consists of adding the scalar \( \rho \) to the value that enters the activation function, as is depicted below. Geometrically, it represents a rotation of the hyperplane present in the latent space of the perceptron. The ARP is a generalization of the classic perceptrons.