Using deep supervised learning algorithms to cluster data

The objective of this project is to develop an algorithm to cluster data by combining geometric-based clustering algorithms and supervised deep learning neural networks and to specify the conditions of the data amenable to this algorithm.

Robust interpretability method for deep learning-based models

Despite its ability in solving difficult tasks, deep learning is still widely considered a black-box technology due to its vagueness in explaining its decision. To clarify such vagueness, several recent studies have focused on developing interpretability methods for deep learning-based models. We will develop an interpretability method that can generate robust explanation from a model when the model is under adversarial attacks. We pay attention to establishing an interpretability-retained training framework.

Toward asymptotic optimal dynamic pricing experimentation policy: multi-armed bandit-based dynamic pricing algorithm design

With the growth of e-commerce platform, how to make a good pricing decision becomes a key issue. We will study the optimal dynamic pricing policy (from real-time data to timely decision) for the online market and design decision intelligence algorithms for it. The algorithm aims to carry out the proposed pricing policy, which can automatically transform real-time data into optimal pricing decisions at each time point.

Facilitating AI medical models with statistical methods: an imaging-genetics joint analysis

This project aims to make use of health big data to advance the current AI models and methods with statistical thinking. With the performance breakthrough of deep neural networks and GPU implementation, AI models have been created for diagnostic classification, lesion segmentation, pathological prediction in the clinics. In this project, we intend to build automated AI systems for disease diagnosis and precision treatment using genetic and medical image data.

Analysis for medical data and medical images in general

In recent months we have been in close contact with a few national medical centers and hospitals. Various research materials such as genetic data, multi-omics data, images, lab tests scores, clinical data and health records were acquired. It is very challenging to integrate the machine learning algorithms with biomedical research, where data often possesses characteristics of being high dimensional, heterogeneous, contaminated, with missing values, censored, biasedly sampled, etc. Sound statistical methods and theory are inevitable for a better understanding and interpretation of the personalized medical diagnosis, treatment and prognosis. This research group will work closely with physicians and is expected to provide clinical suggestions for better healthcare in Taiwan.

Federated learning for disease prediction in three-level organizational structure

Federated learning enables multiple parties to build a common, robust machine learning model without sharing data, thus it allows to address critical issues such as data privacy, data security and access to heterogeneous data. As machine learning has led to innovations in healthcare by its ability to access data, which approximates the true distribution, federated learning is a promising approach to obtain robust statistical inference. The approach has shown impact on the medical treatment pipeline, ranging from improved medical image analysis, providing better diagnostic tools, over true precision medicine by helping to find similar patients, to accelerated drug discovery decreasing cost and time-to-market for pharma companies. This research group aims to practice federated learning in the Veterans General Hospital system in Taiwan, where the unified medical information has deployed in several branches of medical centers, local hospitals and veterans nursing centers. We anticipate improving patient care globally.