An IOS APP that can recognize hand written digits from any materials, using multi-layer perceptron, convolution operation and Canny edge detector.

We collect 4 million pictures and their descriptions from internet, and used 326 labels extracted from those descriptions.

Then we trained CNN and LSTM with them

We achieved amazing results in Chinese image caption and better performance in Flickr database than Google's model

( O. Vinyals, A. Toshev, S. Bengio, and D. Erhan. Show and tell: A neural image caption generator. In CVPR, 2015. ).

See my video at YouTube

I wrote most part of script for this video in Beijing University of Posts and Telecommunications

here you can see the some snapshots in the final video

Full video can be found at YouTube

An image segmentation program based on Gaussian Mixed Model and Expectation Maximization, which can divide picture in to parts according to its color.

Solving N-puzzle problem on your iPhone, based on A* algorithm

An MATLAB program using Principal Component Analysis that can recognize whose face it is in a picture given,

in a very fast way.

I think using specific area rather than weights to store information for neural networks is the future of large data training and hard logical tasks, so I made a program based on Facebook's Memory Networks for further learning.

It conduct better accuracy on certain QA tasks.

I build the first edition of this IOS APP for V-Movie company,

search APP store for "魔力盒" or try it at Magic Box

structure of the Neural Network that generated poems

&

generated results:

《蝶恋花*中国》

(Butterfly loves flower · China)

寒意勒花风卷絮，一曲情怀，已断春归去。不觉山中难醉里，多少林同了处起。

帘卷波回千百转，雁字眉峰，旧日催人度。何似当年后心事，万岁风吹破月见。

This Song Iambics is generated based on its “China” cue phrase.

新春快乐

(Happy New Year)

新为清风老，春来是小朝。快出诗语尽，乐却要如高。

Putting the first word of every sentence above together is "新春快乐" , which means "happy new year". The Acrostic Poems is generated based on this “happy new year” cue phrase.