Why are large 9x9, 7x7, 5x5 matrices almost no longer used?

Large convolutions (e.g., 9x9, 7x7, 5x5) have almost ceased to be used for several reasons:

• High computational cost: large filters require significantly more operations, which slows down training and inference.
• Loss of local information: large kernels cover too large an area, which can lead to blurring important local features.
• Efficiency of small convolutions: using several consecutive convolutions with small kernels (e.g., 3x3) allows models to capture the same receptive field but with less cost and greater non-linearity.

For example, the VGGNet architecture uses several 3x3 convolutions in a row instead of one large one, which improves quality and efficiency.