Neural network based learnings in support of two application domains

Date
2022-10-26
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Artificial intelligence (AI) empowers machines to mimic the behaviors and thoughts of humans. With this technology now being used in all walks of life, it powers many real-world applications, ranging from language understanding to facial recognition. This thesis explores AI applications in two fields: high dynamic range (HDR) imaging and entrepreneurial prediction. In HDR imaging, multi-exposure fusion (MEF) is the easiest way to produce an HDR image without expensive professional cameras. Employing AI in MEF highly enhances the quality of generated HDR images compared to traditional hand-crafted methods. The neural networks (NNs) in these AI applications tend to be complex, with excessive parameters and heavy calculation costs. Hence, we propose a lightweight NN for MEF, consisting mainly of depthwise and pointwise convolution. Experimental results show that this proposed technique could generate HDR images in extremely exposed regions with sufficient details to be legible. Our model outperforms other state-of-the-art approaches in peak signal-to-noise ratio (PSNR) score by 0.9 to 8.7 while achieving 16.7 to 306.2 times parameter reduction. We also successfully develop our lightweight NN model on a Raspberry Pi and a field-programmable gate array (FPGA).In entrepreneurial prediction, this study focuses on female entrepreneurial success enhancement, which is rarely researched by scholars. We collect survey data from the “From Lab to Fulfillment” workshop focused on developing entrepreneurial skills amongst female academics. To augment the collected tabular dataset, we utilize the variational autoencoder (VAE) technique to generate artificial data. After training five machine learning algorithms on the combined dataset of the original and synthetic data, the result manifests that artificial data can improve the training performance of the models. This preliminary result also shows that the NN is the optimal predictive model with 70.3% precision, 74.7% recall, and 79.2% F1-score metric results. Based on this, we retrain the NN model and get a preliminary experimental result that Good Communicator is the most critical factor for female entrepreneurs’ success.
Description
Keywords
Citation
Liu, Z. (2022). Neural network based learnings in support of two application domains (Master thesis). University of Calgary, Calgary, Canada). Retrieved from https://prism.ucalgary.ca .