Sabyasachi Sahoo sabyasachis

	Smaller Filter Sizes	Larger Filter Sizes
	It has a smaller receptive field as it looks at very few pixels at once.	Larger receptive field per layer.
	Highly local features extracted without much image overview.	Quite generic features extracted spread across the image.
	Therefore captures smaller, complex features in the image.	Therefore captures the basic components in the image.
	Amount of information extracted will be vast, maybe useful in later layers.	Amount of information extracted are considerably lesser.
	Slow reduction in the image dimension can make the network deep	Fast reduction in the image dimension makes the network shallow
	Better weight sharing	Poorer weight sharing
	In an extreme scenario, using a 1x1 convolution is like treating each pixel as a useful feature.	Using a image sized filter is equivalent to a fully connected layer.

	Smaller Filter Sizes	Larger Filter Sizes
	Two 3x3 kernels result in an image size reduction by 4	one 5x5 kernel results in same reduction.
	We have used (3x3 + 3x3) = 18 weights.	We used (5x5) = 25 weights.
	So, we get lower no. of weights but more layers.	Higher number of weights but lesser layers.
	Therefore, computationally efficient.	And, this is computationally expensive.
	With more layers, it learns complex, more non-linear features.	With less layers, it learns simpler non linear features.
	With more layers, it necessitates the need for larger memory.	And, it will use less memory for backpropogation.