18
中中中 中中中中中中中中中 -- 中中中中 团团LCLL Zhejiang University

中国云 移动互联网创新大赛 -- 火眼金睛

  • Upload
    cassia

  • View
    114

  • Download
    6

Embed Size (px)

DESCRIPTION

中国云 移动互联网创新大赛 -- 火眼金睛. 团队:LCLL Zhejiang University. Team. Team Leader: Yue Lin ( 林悦 ): responsible for model training, parameters tuning, structure design Team Members: Debing Zhang ( 张德兵 ): responsible for seeking the new techniques in DL, structure design - PowerPoint PPT Presentation

Citation preview

Page 1: 中国云 移动互联网创新大赛 --  火眼金睛

中国云 移动互联网创新大赛 -- 火眼金睛

团队: LCLL

Zhejiang University

Page 2: 中国云 移动互联网创新大赛 --  火眼金睛

Team

Team Leader: Yue Lin ( 林悦 ): responsible for model training, parameters tuning, structure design

Team Members: Debing Zhang ( 张德兵 ): responsible for seeking the new techniques in DL, structure design Cheng Li ( 李成 ): responsible for the images crawler, build the model training environment Xiaoting Zhao ( 赵晓婷 ): responsible for the data labeling

Publications:

Yue Lin, Rong Jin, Deng Cai, Xiaofei He: Random Projection with Filtering for Nearly Duplicate Search. AAAI 2012

Yue Lin, Rong Jin, Deng Cai, Shuicheng Yan, Xuelong Li: Compressed Hashing. CVPR 2013

Bin Xu, Jiajun Bu, Yue Lin, Chun Chen, Xiaofei He, Deng Cai: Harmonious Hashing. IJCAI 2013

Zhongming Jin,Yao Hu,Yue Lin, Debing Zhang, Shiding Lin,Deng Cai, Xuelong Li: Complementary Projection Hashing. ICCV 2013

Yao Hu, Debing Zhang, Jieping Ye, Xuelong Li, Xiaofei He: Fast and Accurate Matrix Completion via Truncated Nuclear Norm Regularization. 

TPAMI 2013

Yao Hu, Debing Zhang, Zhongming Jin, Deng Cai, Xiaofei He: Active Learning Based on Local Representation. IJCAI 2013

Debing Zhang, Genmao Yang, Yao Hu, Zhongming Jin, Deng Cai, Xiaofei He: A Unified Approximate Nearest Neighbor Search Scheme by

Combining Data Structure and Hashing. IJCAI 2013

Debing Zhang, Yao Hu, Jieping Ye, Xuelong Li, Xiaofei He: Matrix completion by Truncated Nuclear Norm Regularization. CVPR 2012

Yao Hu, Debing Zhang, Jun Liu, Jieping Ye, Xiaofei He: Accelerated singular value thresholding for matrix completion. KDD 2012

Zhongming Jin, Cheng Li, Deng Cai, Yue Lin: Densitive Sensitive Hashing. TSMCB 2013

Page 3: 中国云 移动互联网创新大赛 --  火眼金睛

数据

传统做法:

Ref: Pedestrian Detection: An Evaluation of the State of the Art

Page 4: 中国云 移动互联网创新大赛 --  火眼金睛

比赛数据

Offline test result: 0.9820.

We choose Deep Learning

Page 5: 中国云 移动互联网创新大赛 --  火眼金睛

Structure

We follow the structure used in MNIST, ImageNet

Page 6: 中国云 移动互联网创新大赛 --  火眼金睛

More data is good

Negative Data: Caltech 256, Some images selected in VOC.

Caltech 256

VOC

All the data need to be checked. Mislabeled images will hurt the performance.

Page 7: 中国云 移动互联网创新大赛 --  火眼金睛

More data is good

Positive Data: Baidu Shitu

We implement a crawler to send some classical images to the Baidu Shitu and save the results.

save the page -> get the images’ link + another crawler

Page 8: 中国云 移动互联网创新大赛 --  火眼金睛

Training Information

Gray vs. Color:

Resolution:

Maps:

Convolutions:

Color is better.

128x128 is better than 64x64 and 32x32.

More maps is better but cost more time. Finally we choose 64 maps.

More Convolutions is better but cost more time. Finally we choose 5 convolutions.

Page 9: 中国云 移动互联网创新大赛 --  火眼金睛

Training Information

Local response normalization

We use ReLU neuron types f(x)=max(0,x). (Much faster)

“Local response normalization aids generalization.”

Page 10: 中国云 移动互联网创新大赛 --  火眼金睛

Parameters

Layers Type Maps & neurons

0 input 3 maps of 128x128 neurons

1 convolutional 64 maps of 120x120 neurons

2 max-pool 64 maps of 40x40 neurons

3 normalization 64 maps of 40x40 neurons

4 convolutional 64 maps of 40x40 neurons

5 normalization 64 maps of 40x40 neurons

6 max-pool 64 maps of 20x20 neurons

7 convolutional 64 maps of 20x20 neurons

8 normalization 64 maps of 20x20 neurons

9 max-pool 64 maps of 10x10 neurons

10 convolutional 64 maps of 10x10 neurons

11 convolutional 64 maps of 10x10 neurons

12 Fully connected 512 neurons

13 Fully connected 2 neurons

Page 11: 中国云 移动互联网创新大赛 --  火眼金睛

Viewing the Net

Training and test error over time.

Page 12: 中国云 移动互联网创新大赛 --  火眼金睛

Viewing the Net

Page 13: 中国云 移动互联网创新大赛 --  火眼金睛

Viewing the Net

Page 14: 中国云 移动互联网创新大赛 --  火眼金睛

Viewing the Net

Page 15: 中国云 移动互联网创新大赛 --  火眼金睛

Viewing the Net

Page 16: 中国云 移动互联网创新大赛 --  火眼金睛

Discussion

Dropout

Ref: ImageNet Classification with Deep Convolutional Neural Networks

Achieve better performance on ImageNet, MNIST, TIMIT, CIFAR-10. In offline test, the performance is improved from 0.9820 to 0.9826.

Page 17: 中国云 移动互联网创新大赛 --  火眼金睛

Future

1.Why it works?

Theory Extension: Long way to go

2.How it goes?

Distributed Computation , Huge Data

Page 18: 中国云 移动互联网创新大赛 --  火眼金睛

Thank you