Convolutional neural network kernels during training on MNIST dataset скачать в хорошем качестве

Convolutional neural network kernels during training on MNIST dataset

I created a convolutional neural network with Tensorflow and visualized how the filters of each of the convolutional layers change over time. The left most panel shows the input batch for the most recent training step. The digits in green are the ones correctly identified by the network, and the red ones are the digits that it failed to identify (this is just to give an indication of how the network is doing, and what digits it struggles with. I am not using the training set to calculate network average error !). In the next panel you can see the average error of the network over the test set vs the number of training steps taken. The 2 right most panels show the filters of each of the convolution layers. I applied a Gaussian blur filter to the second layer kernels so that it is easier to see in the graph. If you want to know more please visit my blog: https://sigsegv.xyz/