My graduate research topics are probabilistic machine learning and its clinical applications. During my undergraduate studies, I have also engaged in networks analysis and computer vision.
Here is the catalogue of selected research topics. For a complete list, please visit my Google Scholar.
Graduate Research:
Undergraduate Research:
(with Sarang Joshi, Ross Whitaker)
We proposed an adaptive, versatile Bayesian method for functional principal component analysis, BSFDA. This method scales linearly with dataset size, automatically determines the number of components, excels in modeling sparsely observed functional data, and effectively quantifies uncertainty. It demonstrates state-of-the-art performance in dimensionality selection and mean-covariance structure estimation with the synthetic benchmark data from [1]. Its efficiency is further confirmed through simulations with 4D functional data. Additionally, we tested this method with real-world biomedical and meteorological data, including CD4 cell counts and wind speed measurements, ensuring its applicability and robustness in practical scenarios.
Proportion of accurate dimensionality estimations for Scenario 5 in [1] (rank=6, most challenging):
| $N_i$ | $\text{AIC}_{\text{PACE}}^{[1]}$ | $\text{AIC}^{[1]}$ | $\text{BIC}^{[1]}$ | $\text{PC}_{p1}^{[1]}$ | $\text{IC}_{p1}^{[1]}$ | $\text{AIC}_{\text{PACE}}^{2022 [2]}$ | $\text{BIC}_{\text{PACE}}^{2022 [2]}$ | $\text{fpca}^{[3]}$ | $\text{BSFDA}$ |
|---|---|---|---|---|---|---|---|---|---|
| 5 | 0.705 | 0.470 | 0.090 | 0.070 | 0.545 | 0.425 | 0.410 | 0.855 | 0.925 |
| 10 | 0.065 | 0.570 | 0.525 | 0.775 | 0.705 | 0.575 | 0.575 | 0.500 | 1.000 |
| 50 | 0.000 | 0.260 | 0.590 | 0.980 | 0.965 | 0.870 | 0.770 | 0.695 | 1.000 |
Mean squared error of covariance for Scenario 5 in [1]:
| $N_i$ | $\text{AIC}_{\text{PACE}}^{2022 [1]}$ | $\text{BIC}_{\text{PACE}}^{2022[1]}$ | $\text{fpca}^{[3]}$ | $\text{refund.sc}^{[4]}$ | ${\text{BSFDA}}$ |
|---|---|---|---|---|---|
| 5 | 12.373 ± 4.026 | 12.377 ± 4.031 | 5.192 ± 6.166 | 8.833 ± 4.730 | 5.814 ± 3.535 |
| 10 | 10.391 ± 2.521 | 10.391 ± 2.521 | 2.098 ± 1.425 | 5.314 ± 3.501 | 2.068 ± 1.427 |
| 50 | 9.054 ± 1.683 | 9.054 ± 1.683 | 1.642 ± 1.240 | Prohibitive | 1.638 ± 1.247 |
This manuscript is currently being finalized.
Reference:
Proposes a method for modeling high-dimensional medical image data using a Gaussian model that accounts for inherent correlations, integrating a Gaussian-process noise model within a generative framework and optimized through a novel expectation maximization (EM) algorithm, validated on synthetic and anatomical shape data.
Introduces an unsupervised machine-learning approach using the cranial morphology deviation (CMD) metric to objectively quantify the severity of metopic craniosynostosis, demonstrating strong alignment with expert surgical assessments and providing a standardized tool for guiding operative decisions
Demonstrates that 3D photography, combined with a machine learning algorithm, offers a radiation-free, accurate, and objective alternative to CT imaging for quantifying the severity of metopic craniosynostosis, improving patient care and reducing exposure to unnecessary radiation.
Check it out at craniorate.org.
Introduces a machine-learning tool with two unique severity metrics—metopic severity score (MSS) and cranial morphology deviation (CMD)—that quantitatively assess metopic craniosynostosis, providing clinicians with an objective, point-of-care tool for improved decision-making through a cloud-based online portal.
Uses CranioRateTM, a machine learning skull morphology assessment tool, to quantify overcorrection and longitudinal morphological changes in metopic craniosynostosis patients undergoing fronto-orbital advancement (FOA), finding that while patients initially exhibited overcorrection, they approached normal skull morphology two years postoperatively, with more severe preoperative cases showing a greater risk of postoperative regression.
Presents a deep learning model that jointly learns population-level and subject-specific anatomical shapes directly from volumetric images, outperforming traditional methods by providing an efficient and accurate framework for statistical shape modeling and segmentation.
Integrates neural collaborative filtering with a text-aware regularizer to enhance recommendation accuracy, leveraging review texts during training for improved prediction performance while maintaining efficient inference in real-world applications.
In academia, authors from different affiliations publish papers in various conferences/journals. The activities and interactions of these entities generate myriad academic records. Through collecting and analyzing these records, we can obtain lots of interesting findings on the academic system, for instance, to sense the rising topics for research.
This system helps people accurately and easily get the information from immense raw data, to solve the problems with overloaded data. Similar ones include Google Scholar which is widely used and dblp mainly for Computer Science.
My work could be divided into 3 parts: mentorship prediction, recommendation and visualization. You can have a glance about my work at predicted mentorship in author page, predicted mentorship by area and register as a user to see the recommendation in the user zone.
More specifically, I devised and implemented the system to predict links of mentorship using academic records, provide personal mentor recommendations and convey the information to the users with frontend visualization.
The forpart work (for a) include mass data loading, cleaning, processing and learning. In the means of visualization, we constructed maps (force-directed) in the demensions of published years, conferences and topics or fields. We excavated the various relationships among the papers, conferences, topics and authors.
I implemented the data crawling, training and prediction of mentor relationship using artificial neural network for AceMentor, which aimed to implement link prediction for mentor relationship and visualize the network. Following this, I built a mentors recommender system for students mainly based on Collaborative Filtering and content.
Automative driving is the core part of Intelligent Vehicle. We applied the formation of space-time map, base on Spatiotemporal Theory, selected and tested the road points in the map to fit into three-order model. The system is designed to improve the efficiency of lane inspection and the robustness of the fuzzy lane line.
Reliability is that when the lane line is not clear visible, we applied the predicted road points in space-time map of multiple images to solve this problem. Compared with the traditional lane mark detection method, this method is simple for installing the camera, does not need to be ortho-projection transformation, and is not sensitive to the vehicle bumps. However, this method still failed at excessive exposure.
Based on Efficient Lane Detection Based on Spatiotemporal Images.
If you are interested in this work or have any questoin, please contact me via wenzheng.tao@utah.edu.
The copyright belongs to the author. For business reprint please contact the author to obtain authorization, non-commercial reprint please indicate the source.
©2017-2020 by Wenzheng Tao. All rights reserved.