Homework 2

MNIST

Train two different kinds of classifier on all 10 digits of MNIST:

1. Use the high-level API of chapters 1 and 2 to train a ResNet.
2. Use the low-level approach of chapter 4 to train a one-layer MLP; use a 30-dimensional hidden layer with ReLU activations. (This is exactly the network discussed in chapter 4, “Add a Nonlinearity”.)

Just like in Homework 1, describe:

• How confident are you in these conclusions? Why?
• What decisions did you make in the process? Why? What implications did those decisions have?

For reference, here is state of the art on MNIST. How does your accuracy compare?