Hello everyone, for the first part of this lecture's exercises, you will return back to the MNIST dataset to implement a convolutional neural network to recognize handwritten digits. Afterwards, you will explore image classification on the CIFAR-10 dataset. In previous exercises, you classified MNIST images using a multilayer perceptron with fully connected layers. Here, you will instead implement a CNN. We will use the same code as before to set up our model, including setting up the compute backend and loading MNIST dataset. However, when defining the model architecture, you will instead use a convolutional neural network. To do so, add a convolutional layer to the list layers using the compound layer conv with 25 by 5 filters and a padding of 2. Then append a 2 by 2 max-pooling layer, followed by affine and dropout layers, and ultimately an output layer with 10 outputs corresponding to the 10 digits. Don't forget to import the necessary layers, and as always for more information on the various layers in Neon, you can go to the Neon documentation. Now that you've created and tested your first convolutional neural network on the MNIST dataset, you will now explore image classification using the conv net on the CIFAR-10 dataset. As you go through the exercise, you will notice that there are many kind of topologies that can be used in convolutional neural networks. From as simple as a model with one conv layer max Boolean and fully connected layer to a deep residual network. Try out different topologies and see how that affects the model. And also go through and fill in the missing components of the deep residual network implemented in the exercise. Again, if you run into issues, please post in the forum alongside this exercise and someone will help you. [MUSIC]
