Machine Learning Q and AI

APPENDIX: ANSWERS TO THE EXERCISES

Chapter 1

1-1. The final layer before the output layer (the second fully connected layer in this case) may be most useful for embeddings. However, we could also use all other intermediate layers to create embeddings. Since the later layers tend to learn higher-level features, these later layers are typically more semantically meaningful and better suited for different types of tasks, including related classification tasks.

1-2. One of the traditional methods of input representation that is different from embeddings is one-hot encoding, as discussed in Chapter 1. In this method, each categorical variable is represented using a binary vector where only one value is “hot” or active (for instance, set to 1), while ...