Academic Homepage
Incoming PhD Student, University of Alberta
Supervisor: Prof. Witold Pedrycz.
I investigate how prior knowledge can be systematically integrated into machine learning systems, with particular attention to optimization, representation, and granular knowledge structures.
Email:
This website documents ongoing research exploration. All ideas are preliminary and subject to further validation.
Research Interests
Question 01
I am interested in how the same prior knowledge changes once it is introduced through data, the hypothesis set, the learning algorithm, or the final hypothesis.
Question 02
A central comparison is between teacher-student regularization, semantic losses, and declarative constraints, especially in terms of optimization behavior and failure modes.
Question 03
I am exploring whether prior knowledge can be built directly into feature-selection algorithm design, data representation, or the training procedure rather than only to the output stage.
Reproduction Projects
Project 01
A small-scale implementation used to examine how rules shape a teacher distribution and how that distribution is distilled back into a student model.
Project 02
A small-scale implementation used to better understand how logical validity can be translated into probability mass over satisfying worlds.
Project 03
A small-scale implementation used to examine a softer form of structural prior, where global landmarks may regularize sparse local observations.
Current Progress
Next
The next stage is to connect the optimization-side comparison to earlier knowledge entry points, especially granular structures and knowledge-guided feature-selection design, without losing the mechanism-level clarity built so far.
Beginning of May 2026
The toy has been finished and the code has been traced, but the next step is to analyze the optimization behavior and failure modes of the teacher-student mechanism, and to read DL2 for a more direct comparison with semantic loss.
Late April 2026
The site now includes bilingual notes on granular computing for machine learning, fuzzy-to-granular model transitions, and knowledge landmarks, extending the project beyond exact or declarative constraints.
Late April 2026
The repository now contains small-scale reproductions for Logic-Net, Semantic Loss, and Knowledge Landmarks, making it easier to compare teacher shaping, satisfying-world losses, and softer global regularization in code.
March-April 2026
The core line is now anchored in one repeated comparison: how prior knowledge changes once it enters learning through posterior shaping, direct symbolic losses, or declarative constraint handling.
Contact
The main external entry points are the research site itself, the public repository, and the note pages that track my current reading and implementation-oriented understanding.