Cha, Jaehoon
(2020)
Artificial Neural Network Design Approaches to Multi-Channel Information Analysis.
PhD thesis, University of Liverpool.
|
Text
201229201_Oct2020.pdf - Unspecified Download (8MB) | Preview |
Abstract
In recent years, a large amount of multi-channel data has been collected due to advances in technology such as with computers and the Internet. However, obtaining and labelling data are still laborious and time-consuming. Yet another issue that adds to the difficulty is finding important channels and features from multi-channel data since having enough channels alone does not guarantee designing efficient algorithms due to scalability problems. In this thesis, a generative model and hierarchical learning models are introduced to deal with the aforementioned issues. First, the learning process of Variational Autoencoders is analysed. Taking into account the role of the mean and the standard deviation, which are used in the reparameterization trick, we propose a new generative model. The proposed model is modified from the original Autoencoder architecture which is used for dimensionality reduction. The model preserves the architecture of the Autoencoder by removing the reparameterization trick and becomes a generative model by extension of the mapping of the decoder from a discrete latent space to a continuous latent space. The model is compared with VAE and MMD on three benchmark datasets: MNIST, Fashion-MNIST and SVHN datasets. The experimental results show that the difference of the accuracy of the test set when training ANNs using synthetic data generated by the proposed model is less than 10% when training it using the original training set in MNIST and Fashion-MINST datasets. In addition, further experiments are carried out to investigate the impact of the number of the training set when training generative models. The results show that the accuracy of the test set decreases less than 10% when the number of the training set decreases in the NNIST and the Fashion-MNIST dataset. Second, two types of hierarchical learning models are proposed. Designing these models began with the idea of utilizing an innate hierarchy of targets. The first type of model, HAL, is proposed when targets are discrete. This model involves inserting the auxiliary block to output the auxiliary scores from the coarse classes. These scores are distributed based on the corresponding coarse classes. Although the model improves the accuracy of a test set, it has the disadvantage of requiring the coarse classes at the test phase. The second type of models are proposed when targets are continuous. C-FNNs and HADNNs are proposed to perform the regression task by utilizing the coarse classes. C-FNNs and HADNNs are evaluated on three benchmark indoor localization datasets, examples of multi-channel data. Results show that C-FNNs increase the floor accuracy by 30% at least and 60% at most in the three datasets. However, C-FNNs require more than three times the parameters than the baseline. HADNNs achieve better accuracy than C-FNNs and require 1.2 times the parameters than the baseline at most. Third, human motion data is analysed in order to show the importance of the relationship between sensor locations and motion types when identifying motion types. The data were gathered from patients and students in Inha University Hospital, Korea. Twenty-three subjects participated in the experiment and all had to perform nine motion types. Forty-eight total measurements were obtained from eight different body parts. The motion type detection algorithm is divided into five steps and is evaluated based on four metrics: recall, precision, accuracy and F-measure. The proposed detection algorithm has $0.8986$ average recall, 0.9071 average precision, 0.9739 average accuracy and 0.8977 average F-measure. The detection algorithm outperforms PCA, which is a popular method in feature extraction. This shows the importance of feature extraction based on the relationship between channels and targets in multi-channel data. Finally, the motion type detection process is proposed by integrating the proposed models. The process is divided into three: generation, labelling and classification. In generation, the proposed generative model is used to generate synthetic data. In labelling, SVM and PCA are used to label synthetic data. In classification, ResNet with C-FNNs and with HADNNs for a classification task are trained using the combination of the labelled synthetic data and the original training set, and the neural networks are used to detect motion types. The process is evaluated using InhaMotion and nine open source human motion datasets. The results show that training ANNs with synthetic data prevents overfitting, and the proposed generative model outperforms VAE, beta-VAE and MMD. In addition, the combination of ResNet and C-FNNs increase the accuracies of the test sets when coarse classes are available during the training phase. Since C-FNNs do not require coarse classes at the test phase, it is practical to use in daily life problems where hierarchy of targets should be considered.
| Item Type: | Thesis (PhD) |
|---|---|
| Uncontrolled Keywords: | artificial neural networks, multi-channel data, multi-channel analysis, generative neural networks, generative models, hierarchical learning, hierarchical neural networks, human motion detection, autoencoders, indoor localization |
| Divisions: | Faculty of Science and Engineering > School of Electrical Engineering, Electronics and Computer Science |
| Depositing User: | Symplectic Admin |
| Date Deposited: | 07 Dec 2020 16:35 |
| Last Modified: | 18 Jan 2023 23:25 |
| DOI: | 10.17638/03105330 |
| Supervisors: |
|
| URI: | https://livrepository.liverpool.ac.uk/id/eprint/3105330 |
Dimensions
Dimensions
