The focus of this article
Convolutional Neural Network (CNN) is a deep learning model whose main feature is the use of convolutional layers to extract features. The core of the convolution layer is the convolution kernel, also known as the filter (filter). The convolution kernel is a small matrix that extracts features from the input data through convolution operations. How are the parameters of the convolution kernel learned?
The convolution kernel is actually a parameter
We have learned the role of convolution kernel in feature extraction. We know that different convolution kernels extract different features. Scientists have summarized some commonly used convolution kernels through years of research, such as texture extraction. Convolution kernel, a convolution kernel that can detect vertical edges, so in building a convolutional neural network, we don't need to artificially specify what kind of convolution kernel to use to extract features. The convolution kernel can be used as a parameter in backpropagation. Learned, sometimes what are the features extracted by the convolution kernel? We may not understand at all.
As shown in the figure, it is a 3*3 convolution kernel, which has 9 values. Different values determine the difference of the convolution kernel, so the features extracted by the convolution kernel are also different.
In actual use, its value will be updated as a parameter in backpropagation. The parameters of the convolution kernel are obtained through training. Maybe sometimes we don’t know what the trained convolution kernel does. We can understand it. An operation similar to vertical edge detection is done for it, that is, the features of a certain part are extracted, which may be much more complicated than simple edge detection, to detect 45° or 70° or 73°, etc., or even any Angled edges.
Summarize
During the model learning process, more often than not, filters that outperform any of these previously identified functions are automatically learned. Because the neural network will always automatically learn the convolution kernel parameters suitable for the data according to the data. That is to say, when we use the convolution kernel, we don't have to use the nine numbers chosen by those researchers, but should treat the nine numbers in this matrix as nine parameters, and then specifically What are the 9 parameters? It should be determined by the data, and after using the backpropagation algorithm, the goal is to learn these 9 parameters, which determine the function of the filter.