Classic Networks
Last updated
Last updated
It was developed by Yann LeCun et al. in 1998
The goal of this network was to classify handwritten digits
It was trained using 32x32x1 grayscale images
This net is smaller compared to today's standards; it had about 60K parameters
It is interesting to note that as the height and width of the image decreased across the layers, the number of channels increased
Also, the net had a sigmoid non-linearity after the pooling layers, which is no longer used today after pooling
AlexNet was developed by Alex Krizhevsky, Geoffrey Hinton, and Ilya Sutskever in 2012
It is much bigger than LeNet-5 and has about 60M parameters
It was developed by K. Simonyan and A. Zisserman in 2014
Instead of having thousands and thousands of parameters, this model used fixed parameters
All its convolutional layers had 3x3 filters with stride 1 and "same" padding
All its max-pooling layers were 2x2 and had stride 2
It had a very simplified architecture:
