How can I create a non-fully connected neural network?

3 次查看（过去 30 天）
Xiaomin Li
Xiaomin Li 2017-7-18
评论： Xiao Gao 2018-10-16
Can I create a non-fully connected neural network using Matlab? I mean I try to connect the neuron as I want, can I? Thanks!

请先登录，再回答此问题。

采纳的回答

Saurabh Gupta
Saurabh Gupta 2017-7-20
You can create a custom neural network and control the connections as per your requirement. The following documentation describes one example of custom neural network architecture for your reference.
https://www.mathworks.com/help/releases/R2017a/nnet/ug/create-and-train-custom-neural-network-architectures.html
The following documentation page has a wealth of related resources that you can explore and use as applicable.
https://www.mathworks.com/help/releases/R2017a/nnet/define-neural-network-architectures.html
I hope this helps!
  3 个评论
显示 1更早的评论
zafar khan
zafar khan 2018-2-9
HI Li and Saurabh, I am trying to do something similar but i cannot comprehend how to achieve a a sparse network. Although i can change number of neurons in each layer but that doesn not mean network is not fully connected. The weights automatically connect all weights of previousl layer neurons lets 'n' to all of the next layer neuron lets say 'm' creating a m*n weights matrix. How can i avoid this but creating a network where one neuron doesnot receive data for all neurons but from only one. I will appreciate if you can help me with this. I have tried to look in the links but no joy.
Xiao Gao
Xiao Gao 2018-10-16
Hi Khan, I am dealing with the same problem. The strategy I am thinking about is dividing the initial inupt vector into several sub-vectors, even only one variable for one sub-vector, and initializing arbitrarily each sub-vector's connection with the following layers. It seems kind of redundant but I hope it works.

请先登录，再进行评论。

更多回答（0 个）

类别

在 Help Center 和 File Exchange 中查找有关 时序、序列和文本深度学习 的更多信息

标签

产品

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!