Why do we use Convolutional Neural Networks(CNNs) for images/videos? What are the advantages of using CNNs over vanilla Neural Networks?
-
Spatial arrangement: Information in an image makes more sense spatially. CNNs preserve spatial structure in images.
-
Local Connectivity: CNNs take advantage of Local Spatial Coherence in the input. As opposed to vanilla Neural Networks where every neuron is connected to all neuron in the previous volume, in CNNs each neuron is connected to only a local region of the input volume. This local region exposed to a neuron in the input volume is called Local Receptive Field. Receptive field of a single unit within a layer is determined by the size of filter(width, height).
-
Weight/Parameter Sharing: The weights of a kernel or filter used in a layer are learnable. This filter convolves with the input volume to produce an output feature map. The weights of a particular filter is shared across the input which produces this output. Due to this, a feature (say an edge, corner) learned at one spatial location doesn't need to be learned again if the same is found at another spatial location in the input.
Read more:
-
https://cs231n.github.io/convolutional-networks/
-
https://towardsdatascience.com/understand-local-receptive-fields-in-convolutional-neural-networks-f26d700be16c
-
https://towardsdatascience.com/understanding-parameter-sharing-or-weights-replication-within-convolutional-neural-networks-cc26db7b645a
-
https://www.quora.com/What-are-the-advantages-of-a-convolutional-neural-network-CNN-compared-to-a-simple-neural-network-from-the-theoretical-and-practical-perspective
-
https://msail.github.io/post/cnn_human_visual/