Abstract
II
Abstract
With the development of the information technology, Deep Learning based on neural
networks has become the synonym of all kinds of advanced technology. Since the neural network
technology appeared in the last century, so many models based on neural networks have been
used to solve different problems in real world. Especially in 2012, the models based on deep
learning raised the accuracy by 11 percent, in the ImageNet Large Scale Visual Recognition
Challenge (ILSVRC), which led to a new wave of deep learning. Deep learning can learn
effective feature automatically instead of obtaining feature artificially, which can reinforce the
ability of abstract learning, where the deep learning improve the development of image detection,
image recognition, speech recognition, tracking and so on.
Extreme learning machine was proposed in 2004, and in past decades, has made great
strides forward. First of all, many theories have proposed to support the development of the
extreme learning machine; Secondly, extreme learning machine has been extended to real world
to solve real problems. But in extreme learning machine, no matter in weights initialization,
feature representation and in deep learning, there are many optimized problems need to be
solved.
This paper has explored the extreme learning machine and tried to solve three problems:
First of all, this paper has explored the weight initialization of extreme learning machine
and tried to illustrate the effect when different methods are used. The proposed method solves
the problem that the random generated feature is lack of compactness and discriminant, and
provides a fast and effective method to initialize a neural network model.
Secondly, this paper has explored the feature representation of extreme learning. Distributed
representation is a common method in feature representation. Without tuning the weights of
extreme learning machine iteratively, this paper has proposed the distributed representation
based on extreme learning machine. It not only adds the category information into the model by
combining the features, but also improves the performance of extreme learning machine.
Thirdly, this paper has explored the relationship between fully connected neural networks
and convolutional neural networks and proposed the convolutional extreme learning machine
based on distributed representation, and extend extreme learning machine to the deep learning
field in the aspect of convolution. Convolutional learning can learn more local information using
the local receptive fields, which can learn a more represented feature. Extreme learning machine
万方数据
