Hi, you have found my personal website! This is a place where I collect my ideas and review topics primary on machine learning and AI.
My journey on machine learning started when I was in college Jilin University. Back then (year 2002) I was studying for two bachelor’s degrees simultaneously - Computer Science and Biology. It was a weird combination in many people’s eyes, but I really enjoyed both subjects. The courses were taught independently, and it never occurred to me we could study biology with the help of computer science until Prof. Ying Xu, a computer scientist who converted to Bioinformatician give a talk on his work in bioinformatics. His work was heavily on predicting the protein 3D structures - applied machine learning in biochemistry. Well, he never mentioned the words machine learning, because it was not very popular then and we were still in the AI winter. But the seeds was planted.
After college, I joined BGI and there was the tail of the human genome project. Exciting time for bioinformatics - once we have the human genomes stored in the hard disk, that is the next natural step! Jun Wang was the director of research in BGI, he was a star! He sat a few cubics away from me in BGI’s big open plan office, and I admired him. Jun Wang has a background in Neural Networks and he really believed the future of genomcics is in computation. Surrounded by high profile scientists, I realized I needed a PhD.
I started my PhD in University of Luebeck late 2008. I was lucky enough to have Prof. Thomas Martinetz as one of my supervisors. He is one of the first appointed neuroinformatics professors in Germany. Neuronal Networks was still in ice age, Thomas once told me to avoid it and look into SVM and Sparse Coding. To be fair, Sparse Coding and ConvNet are closely related, here is a nice paper, and a good slides.
Anyway, the unpopularity of Machine Learning in 2008 leads to less funding. So in order to survive as a poor researcher, we had to render our research on machine learning applied to a specific field. In my case, my host institute was studying x-ray crystallography on viruses proteins. I tried to marry Machine Learning to it. My research work slowly shifted towards mathematical modeling of biological sequences (DNA, RNA and Protein sequences), I discovered a whole new world of computational phylogenetics.
I moved to a very young and talented research group in Heidelberg Institute for Theoretical Studies headed by Prof. Alexandros Stamatakis. There we fit Markov Processes model to DNA sequences to try understand how nature shapes DNA sequence evolution. We were not only developing the best mathematical models for bioinformatics, but also delivering high quality software implementations, it was fun. But more importantly, the lab was full of nice and talented people, I loved it there.
Phylogenetics is quite far from Machine Learning, well, we can argue it is a sort of unsupervised learning. But my journey on machine learning surely did not end there. Because of my background in modeling DNA sequence data, I got a job at Oxford Nanopore(ONT) after my PhD. ONT was developing a DNA sequencer that can read DNA sequences via a tiny hole - a nanopore. The tech is fascinating, as DNA strand translocate the small hole, an electronic signal can be recored. The size of the hole was designed not to read one DNA base at a time but a couple, we call a k-mer. Then the task was to decode the electronic signals of overlapping k-mers back to the DNA sequences.
It is a sequence to sequence modeling problem similar to speech recognition and named entity recognition - plenty of research and ideas from the machine learning world. Unfortunately, the researchers in ONT did not manage to make the association and they were stuck with an HMM model - well known and popular in the bioinformatics world. The HMM model was a poor fit to ONT’s physical model, attempts to make the HMM model more complex only lead to explosion in computation time and was estimated the computational cost will be much higher than making the sequencer itself. It was a dead end. When I first proposed using a recurrent neural network (RNN), nobody believed me it could work. When I implemented the first version using a simple bidirectional LSTM and presented the results - which increased the sequencing accuracy by almost 100% in relative measure, everyone was shocked and my results were questioned for months. A byproduct of the new RNN model was it also reduced the computational cost by 100X! This was a hugely successful story of Deep Learning in industry, one model change, overnight, saved ONT’s years of research on the sequencer hardware and hundreds of millions research cost. I personally do not believe there will be another AI winter, but if there were, at least this is a proof of all the frenetics in AI was not in vain.
I guess my experience in ONT resurrected my passion in having a career AI, so I moved on to a London based AI company BenevolentAI. At BenevolentAI, we are trying to use AI to accelerate drug discovery, for a glimpse of what we are doing here, check on our research director’s blog.
Finally, what is this website about? AI and Machine Learning can be incredibly useful and powerful, if you are still not convinced from my personal experience:
- Checkout Google Translate, just try it and you will be amazed by the quality of machine translation.
- Try any of the voice assistance on your smart phone, if you were disappointed by their performance before, I am sure you will find they are much much improved now.
- Read the story of AlphaGO
These are all AI successful stories serving as hard proof that AI and Machine Learning are useful. So if you are not using it in your work or search, you should.
However, not everyone has the privilege of having a PhD in machine learning. Many of you are probably like me, you have some related knowledge, the math in AI does not look that scary to you, but many terms are new. You want to use machine learning, but you are not confident that you know enough. Most of the tutorials you found are either too easy, out dated, or not comprehensive to be useful. There are too many papers on Arxiv, you are not sure should you be worried about a paper posted yesterday or should you be really using another highly cited one published 3 years ago.
That said, my goal here is to help you to learn, to use machine learning and AI, and be more confident that you are in the right path. I will write review like articles that are easy to are understand for people without a Machine Learning PhD, but are deep enough to cover the state-of-art. I’ll also share some of my ideas and my personal research here.
So, enjoy reading and welcome to reach out (see the bottom of this page).