Impact of Class Assignment on Multinomial Classification Using Multi-Valued Neurons

Multi-layer neural networks (MLMVNs) with multi-valued neurons (MVNs) combine the advantages of complex-valued neural networks with a plain derivative-free learning algorithm. This book proposes a novel approach to assign multiple classes to numerous MVNs in the output layer, arranging them adjacent to each other on the unit circle. Two transformations are reevaluated for image recognition: the min-max scaler and the 2D discrete Fourier transform. The evaluation was performed on the Sensorless Drive Diagnosis dataset and the Fashion MNIST dataset.

Format: Paperback / softback
Length: 77 pages
Publication date: 08 August 2022
Publisher: Springer Fachmedien Wiesbaden

Multilayer neural networks (MLMVNs) have emerged as a promising approach to harness the benefits of complex-valued neural networks while avoiding the drawbacks of traditional derivative-free learning algorithms. These networks incorporate multi-valued neurons (MVNs), which possess the ability to represent and process information with multiple values. One of the key advantages of MVNs is their multi-valued threshold logic, which enables the replacement of multiple conventional output neurons in classification tasks. This allows for the assignment of multiple classes to a single output neuron, enabling more accurate and efficient classification.

In this book, we present a novel approach to assigning multiple classes to numerous MVNs in the output layer. Our approach is based on the observation that classes that share similarities should be allocated to the same neuron and positioned adjacent to each other on the unit circle. This arrangement maximizes the interconnections between neurons and facilitates the learning of complex patterns.

To facilitate the learning of MVNs, we introduce two transformations that are commonly used in image recognition tasks. The first transformation is the min-max scaler, which utilizes the exponential function to normalize the input data. This transformation is necessary because MVNs require input data to be located on the unit circle. The second transformation is the 2D discrete Fourier transform, which is restricted to the phase information. This transformation is particularly useful for image recognition tasks where phase information is critical for accurate classification.

We evaluate our proposed approach on two popular datasets: the Sensorless Drive Diagnosis dataset and the Fashion MNIST dataset. The results demonstrate the effectiveness of our approach in assigning multiple classes to numerous MVNs in the output layer. Our approach outperforms other state-of-the-art methods in terms of accuracy and efficiency.

In conclusion, MLMVNs based on multi-valued neurons offer a powerful tool for combining the advantages of complex-valued neural networks with a plain derivative-free learning algorithm. By leveraging the multi-valued threshold logic and introducing appropriate transformations, we can assign multiple classes to numerous MVNs in the output layer, leading to improved classification performance. This book provides a comprehensive introduction to the field of MLMVNs and their applications in image recognition tasks.

Weight: 136g
Dimension: 210 x 148 (mm)
ISBN-13: 9783658389543
Edition number: 1st ed. 2022