Miks kasutatakse närvivõrkudes 1x1 tuumaga konvolutsiooni?

Convolution with a 1x1 kernel in neural networks is used for several purposes:

• Changing the number of channels (depth): 1x1 convolution allows effectively reducing or increasing the number of channels without affecting the spatial dimensions of the image.
• Introducing non-linearity and channel mixing: after 1x1 convolution, an activation is usually applied, which helps the model learn more complex representations.
• Optimizing computations: 1x1 convolutions are often used in architectures like ResNet or Inception to reduce computational load, replacing heavier convolutions.

Example: if you have a tensor of size 28x28x256, applying a 1x1 convolution with 64 filters will give an output of 28x28x64, effectively reducing the feature dimension.

Thus, 1x1 convolution is a tool for managing dimensionality and increasing the expressive power of the model without changing the spatial characteristics of the data.