DL-06 - Transfer learning

Question 1

DL-06 - Transfer learning

Describe transfer learning.

Answer 1

Transfer learning is the process of utilizing a pre-trained neural network model, typically on a similar task, and fine-tuning it for a new, specific task.

Question 2

DL-06 - Transfer learning

What are the stages in a transfer learning workflow? (5)

Answer 2

• Remove unnecessary layers.
• Freeze layers.
• Add new, trainable layers.
• Train the new layers.
• (Optional) Fine-tuning.

Question 3

DL-06 - Transfer learning

What does transfer learning depend on?

Answer 3

The size and similarity of your new dataset relative to the old one.

Question 4

DL-06 - Transfer learning

When using transfer learning, there are 4 cases to consider. What are they?

Answer 4

Whether your new dataset is:
- small/large.
- similar/different from the original dataset.

Question 5

DL-06 - Transfer learning

What is the approach when your new data is small and similar to the model’s dataset?

Answer 5

• Replace last layer with your own.
• Freeze everything in the pretrained.
• Train only own layer.

(See image)

Question 6

DL-06 - Transfer learning

What is the approach when your new data is small and different from the model’s dataset?

Answer 6

• Replace some CNN layers and add output FC layer
• Freeze pre-trained.
• Train only own layers.

(See image)

Question 7

DL-06 - Transfer learning

What is the approach when your new data is large and similar to the model’s dataset?

Answer 7

• Keep all layers.
• Replace last layer with your own, maybe test changing last 2?
• Can train whole network.

Enough data to not worry about overfitting.

(See image)

Question 8

DL-06 - Transfer learning

What is the approach when your new data is large and different to the model’s dataset?

Answer 8

• Keep all layers.
• Replace last layer(s) with your own.
• Train everything (Or several later layers).

In this case, the CNN layers are mostly retrained from scratch. But we could as well initialize it with the pre-trained weights.

(See image)

Question 9

DL-06 - Transfer learning

What is the approach when your new data is large and different from the model’s dataset?

Answer 9

• Replace one or more later layers with your own, e.g. last conv + all full layers.
• Train everything.

(See image)