samuel500 / mnist_neural_network_meta

MNIST_Neural_Network_META

Simple fully connected neural network trained using batch gradient descent on the MNIST dataset. Sigmoid activation function.

Meta

"Meta" neural networks train on the activation of the hidden layers of previously trained neural networks. META_NET_l256#59904#EPOCHS(24).nt is only trained on the hidden layers of the 3 other NNs and achieves an accuracy of 0.977 even though the best of the 3 "RESERVE_NN" only has an accuracy of 0.9721.

I plan to do further explore training on the activations of hidden layers of multiple trained NNs using TensorFlow.