SUMMARY

• He has 19+ years project experience about Computer Hardware, Software Development, etc.
• He worked as the CTO for 3+ years.During the past 4+ years Confidential achieved a significant breakthrough in the cutting - edge machine learning, optimization and big data science. As the leading author, he published the original contribution at Computer Science each subfield’s number one flagship venue such as ICML’16, SIGIR’16, SIGKDD’16, TSC’15 (with Stephen S. Yau), and ICML’14, etc. In 2012 Fall semester he taught the Advanced Data Mining course as a lecturer for PhD and master students.
• He supervised PhD students from Hong Kong, etc. In the meanwhile, he gave several invited talks at various well-known research labs in the US. He served as the general chair in DSCI’17, SMMA’17, ’16; as the program chair, or committee member and so on, he reviewed more than hundred papers for AISTATS’17; SDM’17, ’16; KESA’17; NIPS’16; RecSys’16; WWW’16; TKDE’16, ’15, etc.
• Moreover, he led and collaborated with colleagues from the University of Oxford, Harvard University, proposed a number of online prediction and recommendation algorithms in large-scale machine learning including centralized and distributed environment, and deployed the state-of-the-art technologies to the Boeing airplanes all over the world, which has improved and saved billions of people lives.

PROFESSIONAL EXPERIENCE

Member of the Research Staff, Department of Engineering

Confidential

Responsibilities:

• Push forward the state-of-the-art in speech and language processing and machine learning/AI joins the major research project “Augmenting Communication using Environmental Data to drive Language Prediction”
• Deep/Conditional/Variational/Contractive Auto-Encoder
• (Pre-training) Deep/Convolutional Neural Network (DNN/CNN) for Speech Recognition, Computer Vision, Natural Language & Text Processing, Web Search, Ranking, News Feed
• C++/Java/Python/R/Matlab/Scala/Go, Keras, TensorFlow, Torch, Theano
• Restricted Boltzmann Machine, Deep Belief Network, Deep Dream/Style
• Hidden Markov/Gaussian Mixture model, Expectation-maximization
• (Leaky/Parametric) Rectified Linear Unit, MaxOut, Regularization, Dropout
• Mini-batch/Stochastic Gradient Descent, Backpropagation (Through Time), Momentum, Adadelta/Adadelta/Adagrad/RMSprop, (Multi-variable) Taylor Series
• Modularization and End-to-end Learning in Image/Speech/Text/Vision
• (Prediction-based) Word/Document/Semantic/Locally Linear/Multi-lingual/Multi­domain Embedding, Continuous bag of word model, Skip-gram
• Laplacian Eigenmaps, T-distributed Stochastic Neighbor Embedding
• Semi-supervised Learning for Generative Model under Low-density Separation, and Smoothness Assumption, Transductive/Inductive learning, Self-training
• (Pixel/Bi-directional/peephole/multi-layer/Clockwise/Structurally Constrained) Recurrent Neural Network (RNN), (Conditional) Generative Adversarial Network, Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU)
• Distributed Representation, (Probabilistic/Kernel) Principle component analysis, (Probability) latent semantic analysis, Linear Discriminant Analysis,

Multidimensional Scaling

Confidential

Responsibilities:

• (Layer) Transfer Learning for Speech and Image, Multi-task Learning and Progressive Neural Networks
• Domain-adversarial training and gradient reversal layer, Zero-shot learning
• Sigmoid Kernel and one hidden layer Neural Net for Audio Signal Segment
• Structured Learning for Speech Recognition, Signal Processing, Image Processing, Machine Translation, Syntactic Parsing, Object Detection, Summarization, Retrieval
• Sequence Labelling for POS tagging, Conditional Random Field, Connectionist Temporal Classification, Sequence to Sequence Learning, Energy/Attention- based Model, Neural Turing Machine, Chat-Bot, Video/Image Caption Generation, Visual/Speech Question Answering
• Bagging, Decision Tree/Random Forest, Boosting, (Large Margin) Adaboost, Gradient Boosting, Stacking
• Asynchronous Advantage Actor-Critic (A3C)
• Our project develops ambitious and novel assistive technology for non­speaking people. Specifically, it aims to increase the communication rate and improve the communication experience of people who use Voice Output Communication Aids by developing a system, which leverages multimodal contextual data to inform a probabilistic language module.
• The role presents an opportunity to make major inroads into a novel application area of recent speech and language processing and deep learning progress

Research Scholar

Confidential, Silicon Valley, California, US

Responsibilities:

• Matrix factorization has been studied traditionally by the machine learning community and has many practical applications, such as collaborative filtering. The main focus of the existing approaches is on estimating the parameters of the matrix. A multi-armed bandit is a popular and general model for learning to act, which captures the fundamental tradeoffs of exploration and exploitation
• In this work, we propose the first online algorithms that can find the maximum entry of a rank-k matrix in a sample efficient manner. The algorithms are practical and can solve various real-world problems, such as finding the best campaign-segment pair in Adobe Campaign.
• In many application domains such as artificial intelligence, information retrieval, recommender systems and so on, the classification is used a single or intermediate step towards estimating the class label prevalence. Our goal is not only to estimate whether a specific Flight Deck Effect conveys positive or negative signal but also what is the overall distribution of examples during an event time window, which has been referred to as the quantification in literature
• We propose the first online bandit algorithms for two clusters of models including classification and quantification, we provide the tight regret bounds. Through a number of real benchmark corpus and data sets, we showcase the comprehensive experimental results which outstandingly beats a bunch of state-of-the-art data mining and machine learning methods
• 1) Simple online bandit methods in the stochastic setting Sharp theoretical guarantee compared to recent advances Small number of passes over data Large-scale big data: does not fit in memory; data comes in as a stream Efficient and computational cheap updates Make the best use of data seen so far Deployed to the Boeing airplanes in the world

Technical Consultant

Confidential

Responsibilities:

• EU Project: CrowdRec, where we focused on modeling multi-armed bandit problems that hinge on either online data-dependent clustering or online clustering at both the user and the action side (a kind of co-clustering). The resulting algorithms have been able to significantly increase the CTR (Click-Through Rate) achieved by state-of-the-art bandit recommender system methods on data owned by Confidential (specifically, data coming from user interactions within the Tuenti social network) while retaining the scalability and the adaptivity of standard bandit methods
• Research and development for web application about price comparison and recommendation engine “Shoppydoo”. The application was developed by our team that followed extreme programming (and agile) principles and techniques like TDD, pair programming, code collective ownership, continuous integration, fast feedback, systematic design, etc. Collaborate with colleagues in Spain to develop this project by the following technologies: Ruby (using Sinatra web framework and ActiveRecord for ORM), Postgres, Redis, Nginx, HaProxy, Zabbix for monitoring
• State Key Laboratory of Intelligent Technology and Systems
• Tsinghua National Laboratory for Information Science and Technology
• Cooperated with other students of our machine learning group within national AI lab
• Took participate in group meeting and research presentations and gained experience
• Attended many AI seminars and went deep with large-scale statistical machine learning

Senior Scientis

Confidential

Responsibilities:

• Object Detection and Tracking in Real-Time
• Efficient Image Search Engine for e-Commerce
• Bayesian Network Inference for Genetic Analysis
• Recommender Systems for Personalized Medicine
• Wearable Computers for ECG/EEG Bio-Signal Processing
• Large-Scale Question and Answer Enterprise Platform Service
• Monte-Carlo localization, Kalman filter, Particle filter algorithms
• High availability Linux, Linux Cluster, Heartbeat, Pacemaker, DRBD
• SNMP protocol related software development, Carrier Ethernet GMPLS
• Develop the High Availability function for ZebOS combining with Enea Element Middleware
• Network Card Driver Development; Profibus Porting & Customizing; state diagram, module of DDLM etc
• Wi-Fi module: Marvell 8787 is one of the most promising WLAN chips supporting IEEE 802.11n standard. Our driver is confined to work with only Murata’s modules by Murata’s ID and MAC address in EEPROM
• OS: Support Linux / Google’s Android system standard wireless features
• Platform: Ported and verified on a variety of the host platforms, such as Goshawk (Designed by SyChip. Marvell PXA310 based), Devkit8000 (TI OMPA3530 based), and Harmony (nVIDIA Tegra2 based). Marvell’s
• Independent software: Marvell introduced the new OS independent software architecture for 87xx driver which is totally different with former 86xx Linux drivers. On the other hand, the new 8787 driver was not ready to develop, so we improved the stability and reliability
• Referenced beagle board, 0xlab, rowboat project ported xloader, uboot to the board of panda board, Ported 802.11a/b/g/n WLAN Linux based driver to the Android 2.2 in the board of devkit8000, modified driver that added the protect of accordingly hardware module, replaced the process of load firmware to manual read
• According to the requirement of customer added the display information of region Code, Mac Address, TX, RX packages etc, and accordingly low-level program modified, output related detail design docs and codes
• Participated design review in the US, finished the work of implementation, coordinated with colleague and the third party, the maintenance of the Android Wi-Fi module, wrote the documents of requirement analysis, finished development of test plan and test case, debug the issue of procedure, made optimization of throughput, power save in STA, soft AP mode