The document proposes a novel hand posture recognition method based on curriculum learning with deep convolutional neural networks. The key ideas are:
1) Train a DCNN using curriculum learning with two heterogeneous tasks - segmentation and classification - to transfer knowledge between the tasks.
2) The network is first trained for segmentation task and then transferred to classification task, updating the parameters.
3) Experiments show the proposed method achieves better hand shape classification accuracy compared to training without curriculum learning.
4. 㼀㼛㻌㼟㼛㼘㼢㼑㻌㼠㼔㼕㼟㻌㼑㼝㼡㼍㼠㼕㼛㼚
It requires fundamental knowledge of math that studies along the
curriculum (with other knowledge form different classes)
1
x2 +1
∫ dx
arithmetic equation
differential
square integration
root
psychics ………..
5. 㻷㼑㼥㻌㼕㼐㼑㼍㻌㼛㼒㻌㼠㼔㼕㼟㻌㼜㼞㼑㼟㼑㼚㼠㼍㼠㼕㼛㼚
Inspired from human’s knowledge acquisition
! Train good feature representation using
the curriculum learning
! Transfer the knowledge (networks) from
heterogeneous task
6. 㻰㼑㼑㼜㻌㻯㼛㼚㼢㻚㻌㻺㼑㼠㼟
! Deep architecture which consists of convolutional, sampling and fully
connection layers [LeCun 1998]
! It has translation invariance of object
! CNN+ ReLu, dropout, Normalization, etc [Krizhevsky 2012]
! Recognize the category of 1000 classes
! Top performance in Large Scale Visual Recognition Challenge 2012
7. 㻯㼡㼞㼞㼕㼏㼡㼘㼡㼙㻌㻸㼑㼍㼞㼚㼕㼚㼓
! Train while changing difficulty of training dataset
䚷䚷(= similar with Bootstrap, but different…)
x1
x2
x3
xi
y1
y2
h1h2yj
y1
y2
y3
hjinitial training with simple set (square size)
update with complexity set (various aspect ratio)
We propose the novel curriculum learning which updates the
network from the heterogeneous task
Y. Bengio, J. Louradour, R. Collobert, J. Weston, “Curriculum Learning”, ICML2009.
8. 㼜㼞㼛㼜㼛㼟㼑㼐㻌㼙㼑㼠㼔㼛㼐
• Train good feature representation using curriculum learning
• Transfer the knowledge from heterogeneous task
hand gesture recognition
main idea of proposed method
-We train the network with two curriculum
-Two curriculums are heterogeneous
9. 㻼㼞㼛㼜㼛㼟㼑㼐㻌㼙㼑㼠㼔㼛㼐䚷ࠥ㼠㼞㼍㼕㼚㼕㼚㼓㻔㻝㻕ࠥ
Train the networks as segmentation task
Convolutional Layer Pooling Layer fully connection Layer
Convolutional Layer Pooling Layer
Binarization layer
Input data : gray scale image
ground truth : hand segmented image
10. 㻼㼞㼛㼜㼛㼟㼑㼐㻌㼙㼑㼠㼔㼛㼐䚷ࠥ㼠㼞㼍㼕㼚㼕㼚㼓㻔㻞㻕ࠥ
Transfer the networks to classification task
Utilize as initial parameters
Input data : gray scale image
ground truth : class label
updating the parameters
22. without curriculum learning with curriculum learning
classification class
Ground Truth class
classification class
23. 㼂㼕㼟㼡㼍㼘㼕㼦㼍㼠㼕㼛㼚㻌㼛㼒㻌㼗㼑㼞㼚㼑㼘㼟
1st
convolutional
layer
2nd
convolutional
layer
without Curriculum learning with Curriculum learning
total updating time : 200000 total updating time : 200000
(segmentation: 50000 +recognition:15000)
30. 㻯㼛㼚㼏㼘㼡㼟㼕㼛㼚
! We propose the training method of Deep Convolutional
Neural Networks with curriculum learning
! As the curriculum, the method transfer the network from
heterogeneous task (segmentation = classification)
! The method is able to improve the feature representation
! Future works
䚷䚷apply to other objects and new curriculum