Algorithms benchmarked to the MNIST dataset
Simple fully-connected network
MNIST accuracy: 96%
Simple convolutional neural network
MNIST accuracy: 98%
Notes: Slow, only ran for 2 epoch.
Implementation of the VGG16 architecture with pretrained MNIST data
cite: https://www.cs.toronto.edu/~frossard/post/vgg16/
Note: Because MNIST is 28x28, first 10 layers omitted and layer depth made a factor of 3 smaller (as MNIST has only 1 color layer)
MNIST accuracy:
A symmetric autoencoder for unsupervised learning
Notes: Success is largely dependent on layer sizes, seems to have a small window of convergence
A symmetric convolutional autoencoder
A Deep Convolutional GAN
TODO: Add in additional hidden layers
MNIST accuracy: 98%
epochs: 100
MNIST accuracy: 95%
epochs: 10