The degree in design is, with sincere gratitude, obtained through the International Design Institute of Zhejiang University and the Studio of Design and Innovation program. My master thesis (with Martin Ester) focus on building interpretable models by infusing deep learning with the crispness of logic, which is related to the general theme of neuro-symbolic integration.

I was born at Lanxi (蘭谿, 'the orchid river'), a typical southern Yangtze river town. My father practices the art of traditional Chinese calligraphy as profession, specialized in Wei dynasty inscriptions, and is also a provincially recognized calligraphy mentor. I received intensive training from him when I was younger, until I reached some local/provincial/national awards and exhibitions and then ceased training, for various reasons. I specializes in the classicals of Zhao Mengfu (趙孟頫) and Chu Suiliang (褚遂良), in particular 妙嚴寺記 and 陰符經.

I start to think about interpretability with analysing and interpreting complex network systems, biological systems and artificial machine learning ones. What I find is that interpretability of a complex system is often unrealistic in the common sense and we lack any proper measure, unless we have a clear mechanistic understanding on what kind of interpretation we need it to be. And this is also the case if we wish to develop and interpret artificial systems, as the bitter lesson keeps telling us the simple rule that general methods which leverage massive computation and parameters is ultimately most effective, and how can we even interpret such large-scale models?

Part of me tries to believe in Occam's razor and mies van der rohe's less is more (partly due to my undergraduate education in design), but I soon realize more is different is inevitable when it comes to complex systems. David Marr's level-of-analysis tells something but not really, in particular it does not guides us if we are making the artificial systems when we merely need some sense of control of "what is going on here".

So what I do in my master thesis is that, if we intentionally assume the separation and combination of perception and reasoning, then we can provide a hybrid computation model consisting of neural networks for perception processing raw data into learned representation and logic program for reasoning on the perceived representation, which employs simple learning rules (thus following the bitter lesson) while still provides us a sense-making interpretations about its decision making.
However, I find more problems when finishing this project compared with when I started. To develop a method that can reliably optimize program is one thing. And besides it, there seems to be a even more important problem of symbolic grounding.