CNN

from 東大1S情報α:

• Instead of just performing “matrix and vector multiplication,” this neural network also processes Spatial Filter.
  • By applying a filter that converts small vectors into scalars to various parts (kernels) of a larger vector (original image), a slightly smaller vector is obtained.
    • It’s kind of like that (blu3mo).
    • This operation is indeed similar to “convolution”.
  • By doing this, it extracts necessary features from the image and eliminates unnecessary information.
    • The method of extraction is naturally trained as well.
• In addition, it is common to perform a process called Pooling after convolution.
  • This is a simple process that reduces the resolution of the image.

from#udacity_intro_to_deep_learning_with_pytorch:

• The key is the Spatial Filter explained in the section on Contour Detection.

• CNN automatically learns the filters.

• The number of convolutional layers is the number of kernels that automatically generate multiple filters.

  • A certain filter may be able to detect dog ears.
  • Another filter may detect dog eyes.

• Sometimes a pooling layer is inserted in between.

  • This is to reduce the size while preserving as much information as possible.
  • There are methods such as taking the average or taking the maximum value.
    • The maximum value method is suitable for image detection as it emphasizes distinctive parts.

• CNN increases the depth of the image while decreasing the width and height.

  • The first layer has a depth of 3 (in the case of RGB), so the width and height are significantly larger.
  • The Convolutional Layer increases the depth, while the Pooling Layer decreases the width and height.