Machine Learning Applications with interpretability: From Medical Science, Object Recognition to Forensic Surveillance

Speaker's Bio:

I just finished my MSc study at the school of computing, Newcastle University, where I learned data sciencemodules, e.g. Machine Learning, Big Data Analytics, Cloud Computing, etc. Out of these, I am especially keen on machine learning. I joined Open Lab last October working on machine learning related projects. In Open Lab, I completed 3 major projects [1,2,3], which involve machine learning methodologies, such as multiple instance learning (MIL), zero-shot learning (ZSL), classifier ensemble and generative adversarial network (GAN), etc. In terms of applications, these works are on the fields of wearable computing for health assessment, object recognition (with attributes) and human identification (biometrics).

Previously, I held another MSc degree in bioinformatics from the University of Edinburgh. Before coming to the UK, I took my BSc degree from Ocean University of China and had 1-year work experience as research assistant in Huazhong University of Science and Technology (Top 10 university in China).

Talk Abstract: 

I will introduce my three machine learning projects. The abstracts are as follows: The first project is to utilize only lightweight wearable accelerometers to achieve objective and reliable General Movement Assessment for perinatal stroke detection. We developed a novel framework ‘Discriminative Pattern Discovery’ (DPD) based on multiple instance learning (MIL) that can strengthen the subtle signal of abnormal movements. Our method outperforms basic level of Prechtl’s standard. 

In the second project, we developed a new approach ‘Dynamic Attribute Feature Ensemble’ (DAFE) for zero-shot learning which eliminates insensitive sample, irrelevant features and absent attributes by dynamically select effective features, attributes and samples in test stage. This method significantly exceeds the state-of-the-art in two popular benchmark datasets. 

The third project is to achieve cross-view gait recognition tasks on 10,000 people using the proposed Discriminant Gait Generative Adversarial Network (DiGGAN) framework. This framework can also reconstruct the gait templates in all views which may serve as important evidences for forensic applications. Our method outperforms other state-of-the-art algorithms significantly on various cross-view gait identification scenarios (e.g., cooperative/uncooperative mode). Overall, these three real-world machine learning applications involve various fields, from medical science to forensics, from wearable computing to computer vision, from small data to big data. Moreover, these applications may require certain levels of interpretability.

1. Health assessment[1]: DPD searches windows/frames that contribute the most to the positive trials. The medical  experts can trace the corresponding synchronized video clips for better understanding.

2. Zero-shot learning [2]: Attributes help to interpret the classification results.

3. ForensicSurveillance[3]: DiGGAN can generate gait templates in all views serving as evidences for forensic surveillance. 

Reference

1. Yan Gao, Yang Long, Yu Guan, Anna Basu, Thomas Ploetz. Towards Reliable General Movement Assessment for Infants using Light-weight Accelerometers. ACM Interactive, Mobile, Wearable and Ubiquitous Technologies (IMWUT). ---Major Revision.

  2. Yan Gao, Yang Long, Yu Guan. Nothing to Learn: Non-parametric Dynamic Attribute Feature Ensemble for Zero-shot Learning. IEEE Winter Conference on Applications of Computer Vision 2019 (WACV) ---Under Review.

  3. Yan Gao, BingZhang Hu, Yu Guan, Yang Long, Nicholas Lane, Thomas Ploetz. Robust Cross-View Gait Identification with Evidence: A Discriminant Gait GAN (DiGGAN) Approach on 10000 People. 33rd AAAI Conference on Artificial Intelligence (AAAI 2019)---Under Review.