In CNN(convolution neural network), can the feature map obtained determinately by a random initialization convolution kernel? if not, how to decide the weights in convolution kernel to obtain the feature maps we need? By trial and rerror, are we shotting if our eyes closed?
Hello Yaozhi Jiang,
A few months back, my supervisor asked me the same question: how is the kernel initialized in a CNN, and what are the values initialized in a kernel?
So after some research I got to know that, we can initialize anything in particular to a kernel, it can be a identity matrix, all ones, a uniform matrix or so on.
As per your question, if you are using TensorFlow, the official TensorFlow documentation mentions that the default initializer they use is Glorot Uniform. You can refer to the official TensorFlow documentation. Also, many other initializers are mentioned there.
https://www.tensorflow.org/api_docs/python/tf/keras/initializers/GlorotUniform
Thanksw to Santosh kumar Panda,
I am puzzled about how to understand the action of weight in convolution kernel?
These weights will be updated and optimized while the trainning. After trainning we can obtain a trainned convolution kenel weights, that we call as last-convolution-kernel.
But for the difference between initialized-convolution-kernel and last-convolution-kernel, do we destroy the difference when the starting of trainning, or not?
With another word, can we design a convolution kernel based on the feature we needed to obtain at the starting of trainning?
you cannot directly control the feature maps by setting the weights in the kernel beforehand. It's a learning process.
to Engr. Tufail,
The textbook answer.
so, but there are some convolution kernels which has weights set by programmer not by machine, in which the weights are pre-setting (initialization) before the learning. for example, the convolution kernel for edge-detection. why?
Yaozhi Jiang In CNNs, feature maps are not obtained randomly by the convolution kernel. The weights in the convolution kernel are initially set, often randomly, but they are adjusted during the training process to minimize the loss function and improve the network's performance. This means that while the initial weights may be random, the final weights are the result of a learning process guided by the data.
adaptive convolution kernel based on input-data distribution, possible or impossible?
- Badges
- Science topic