Research

Most of my recent research projects are about applying Machine Learning algorithms to several tasks such as object recognition, computer generated jazz melody improvisation. Previously, I have ever conducted researches on computer network and web services.

1. Object Recognition

Scatter Component Analysis (SCA)

.....

Multi-task Autoencoders (MTAE)

......

Domain Adaptive Neural Networks

We propose a variant of Neural Networks to perform a domain adaptation task. The problem is to recognize target objects that come from a different domain from the training examples, i.e., learning your "home mugs" images, can computers recognize the "office mugs" ? We embed the Maximum Mean Discrepancy (MMD) metric as a regularization in the NN training such that the hidden layer activation difference between the source and target objects is minimized. This model called Domain Adaptive Neural Networks (DaNN) can provide considerable performance improvements on the Office dataset compared to some recent methods.

Deep Hybrid Networks

We deal with a problem of recognizing noisy handwritten digits over various types of noise, where the information about the noise is completely unseen during training. By a specific combination of a sparse Auto-encoder and stacked restricted Boltzmann machines (RBMs), we show that it can provide a certain level of noise robustness with using only clean images as training examples. We refer to this model as Deep Hybrid Network (DHN).

JazzML

We develop a computer generated jazz melody improvisation system based on k-NN and Neural Networks. This system can generate jazz-like improvisation melody patterns given a sequence of chords. The effectiveness of our system was evaluated by a simple Turing Test.

Can you guess whether these following melodies are played by human or generated by our machine?