As part of the collaboration between PeerJ and the International Work-Conference on Artificial Neural Networks (IWANN), we recently gave an award for the Best Early Career Researcher Presentation at IWANN 2021. The award was won by PhD student Shih-Kai Hung for his presentation of “Facial Image Augmentation from Sparse Line Features Using Small
Training Data.” We recently caught up with Shih-Kai to ask about his research.

If you are organizing a conference or workshop and would like to offer a PeerJ Award at your event, please let us know – communities@peerj.com

*****

Shih-Kai Hung PhD Candidate at The University of Essex, UK. 

Can you tell us a bit about yourself and your research interests?

My name is Shih-Kai Hung, a third-year PhD candidate in the School of Computer Science and Electronic Engineering at University of Essex. I decided to return to university to learn the state-of-the-art technologies of artificial intelligence and machine learning after having been an electric engineer for more than 8 years. I am currently conducting research on deep learning methods for image data augmentation based on small training datasets, under the supervision of Prof. John Gan. A large amount of high-quality training data is critical in machine learning, but it is rare and expensive to be collected in many applications. My research aims to promote the effectiveness of machine learning from small training datasets by generating high-quality synthetic training data of good diversity through data augmentation by developing new generative adversarial networks (GANs).

What originally drew you to this topic?

In many realistic applications of machine learning with small training datasets, various issues have been raised, such as overfitting, over-optimization, mode collapse and so on, as the effectiveness of machine learning heavily depends on the quantity and quality of training data. Generative adversarial networks have emerged in recent years as a powerful deep learning scheme and found interesting applications in image synthesis. It seemed to me that GANs can be a very powerful tool for image data augmentation that would be advantageous over traditional techniques for data augmentation and there is still room for developing novel GAN architectures for generating images with controllable features and diversity. I was motivated to use GANs for improving the effectiveness of machine learning when only limited traning data is available.

What aspect of this research did you present at IWANN 2021?

The paper I presented at IWANN 2021 is entitled “Facial Image Augmentation from Sparse Line Features Using Small Training Data”, which proposes a new conditional GAN for generating high-quality facial images based on a very small set of training images, with sparse line/edge features as conditional input to introduce desirable diversity in the augmented facial images. I expect that the proposed method in this paper would find more interesting applications, which will be explored in my future research.

How will you continue to build on this research?

or my next steps, I plan to develop new models and further improve our proposed methods. I will also extend the experimental results to provide convincing evidence that our proposed methods for image data augmentation are able to significantly improve the classification accuracy of machine learning models in the situation where only a very small training dataset is available.