Ji FENG

(冯霁)

PhD candidate, LAMDA group

Department of Computer Science

National Key Laboratory for Novel Software Technology

Nanjing University

Social links: sina weibo zhihu twitter github

Google scholar: My Google Scholar

Email: fengj_at_lamda.nju.edu.cn

Biography




                  Ji FENG

" ...because the people who are crazy enough to think they can change the world,

are the ones who do."



I joined the department as a PhD student of Department of Computer Science and Technology in Nanjing University since 2015 and became a member of LAMDA Group , led by Professor Zhi-Hua Zhou.

Before that, I obtained B.Eng. Degree in Computer Science from Taishan College , Shandong University . (Taishan College is the so-called "pilot class" led by MOE which intakes only 15 students from 10,000 freshmen per year.) My research interest in SDU was mainly on Information Visualization and Visual Analytics. I also did some fun stuffs in differential privacy and computational geometry. Upon graduation at SDU, I was enrolled in the direct PhD program at Nanjing University.

I was a semi-professional young writer for some time. For instance, I had won some quite prestigious nation-wide awards for young writers including this one (first prize). I abandoned this superpower for quite a while and I do NOT have any intention to continue this path. However, I still do enjoy reading great books and explaining ideas (scientific ideas, mostly) in the simplest way.

Research Interest

"Trees are computer scientists' best friends."

---Donald Knuth



My current research interest includes the following:

  • Deep Forest (a.k.a. tree based deep learning)
  • High Performance Computing (from an AI point of view)
  • Deep Neural Networks
  • Quantitative Trading using AI methods

Publications

"I have discovered a truly marvelous proof of this,

which this margin is too narrow to contain."

--Pierre de Fermat



 

Multi-Layered Gradient Boosting Decision Trees
J. Feng, Y. Yu, and Z.-H. Zhou.
arXiv preprint: 1806.00007.

Multi-layered representation is believed to be the key ingredient of deep neural networks especially in cognitive tasks like computer vision. In this work, we propose the multi-layered GBDT forest (mGBDTs), with an explicit emphasis on exploring the ability to learn hierarchical distributed representations by stacking several layers of regression GBDTs as its building block. The model can be jointly trained by a variant of target propagation across layers, without the need to derive back-propagation nor differentiability.

 

Distributed Deep Forest and its Application to Automatic Detection of Cash-out Fraud
Y-L Zhang, J. Zhou, W.H. Zheng, J. Feng, L.F Li, Z.Q. Liu, M. Li, Z.Q. Zhang, C.C. Chen, X.L Li, and Z.-H. Zhou.
arXiv preprint: 1805.04234.

This is a joint work with Ant Financial. We implemented a distributed version of deep forest and successfully applied it on the detection of cash-out fraud with more than 100 millions of training samples. An easy-to-use GUI was also deployed inside Ant Financial and Alibaba. Now data scientists can build the deep forest model with only few mouse clicked.

 

AutoEncoder by Forest
J. Feng and Z.-H. Zhou.
In: Proceedings of the 32nd AAAI Conference on Artificial Intelligence (AAAI'18), New Orleans, Lousiana, USA, 2018.

We proposed the first forest based auto-encoder by enabling tree ensembles to perform autoencoding tasks. The key motivation is that can we extract useful information along decision paths(rather than the information stored in the leaf node)? Our idea is that by traversing from leaf nodes back to the root and caculating the maximum compatible rules, we can then get an amazingly good estimate of the input pattern.
[code on github]

 

Deep forest: Towards an alternative to deep neural networks.
Z.-H. Zhou and J. Feng.
In: Proceedings of the 26th International Joint Conference on Artificial Intelligence (IJCAI'17), Melbourne, Australia, 2017, pp.3553-3559.

Deep learning models were usually achieved by constructing multi-layered differentiable modules and been trained via back-propagation. In this work, we proposed the first deep model called gcForest via non-differentiable components without the need for backprop. Compared with deep neural networks, gcForest achieves highly competitive performance with much fewer hyper-parameters.
[code on github]

 

Deep MIML Network.
J. Feng and Z.-H. Zhou.
In: Proceedings of the 31st AAAI Conference on Artificial Intelligence (AAAI'17), San Francisco, CA, 2017,pp.1884-1890.

We proposed an unified DNN framework called DeepMIML network for Multi-Instance Multi-Label learning. By introducing the sub-concept tensor layer, the network is able to explore the instance-label relationships and can automatically detect key instances for a particular label.
[code on github]

Work Experience

"If you are good at something, never do it for free."

----The Dark Knight



Quant Intern at WizardQuant

Jan 2015 - May 2015

NOTE: This is the work BEFORE my enrollment at LAMDA

I worked as a "quant researcher" at WizardQuant, a private hedge fund.
My primary contribution is introducing machine learning techniques into High Frequency Trade algorithms.

P.S. I still do quant trading using AI systems built by myself in spare time. Automated trading via AI methods (instead of traditional technical methods) should be the only way to go.

P.P.S. My system works.

Awards

"Winner Winner, Chicken Dinner"

---Anonymous Quotes



Most Bussiness Potential Award, 2017

National CCF-Intel Parallel Computing Competition

This is a contest on implementing high performance computing softwares. With over 400 participants, this award goes to only 3 teams.

Frist Prize, 2013

Outstanding Undergraduate Research Programme

This is a research grand (¥10,000 per project) organized by MOE to encourage undergraduate students to participant in academic research. My research on differential privacy won the first prize.

Grand Prize, 2013

National Entrepreneurship Chanllenge

The National Entrepreneurship Chanllenge held by MOE, is the biggest bussiness plan competition for students in China. I won the grand prize(特等奖) in province.

Bronze Medal, 2012

ACM/ICPC Asia, Changsha Site

This is a contest on programming and problem solving for college students across the world.

Contact Information

"Send me the raven."

---Game of Thrones



Mailing Address:

Ji FENG
Room 912
National Key Laboratory for Novel Software Technology
Nanjing University
Nanjing 210093, China

http://lamda.nju.edu.cn/fengj