Notes from Week 4 of auditing the course¶

Multi-class learning - moving from binary (2 class)classification problem, to 3 possible classes.

Fashion MNIST was an example of a 10 class classification problem

Computer graphics can be used to generate image data for training. We see this in the rock, paper, scissors classification exercise.

Key changes compared to binary classification¶

For multi-class classification, the class mode has be set to be "categorical"
In the model definition - for binary classification, the output layer used a sigmoid activation fucntion which outputs a value close to 1 for one value and close to zero for the other. For multi-class classification, the the output later will need a softmax activation function. Softmax will turn the outputs in the last layer into probabilities that sum to one.
For binary classification, when compiling the model, we used the binary cross entropy loss fucntion. For multi-class classification, we will use the categorical_cross entropy (or spare cross entropy)