Week 3 - supervised learning: Classification

Question 1

Artificial neural networks (ANNs)
- 3 kenmerken

Answer 1

• Network of interconnected units.
• Connections have a certain weights.
• In a feedforward network, activation
  propagates from input units to output units
  through hidden units.

Question 2

Classification

Answer 2

Assigning a discrete label to an input.

Question 3

How do you determine the weights in a ANN?

Answer 3

Supervised learning!
-> Backpropagation of error

Question 4

Backpropagation of error
- 10 steps

Answer 4

1. Initialize the Network:

Start with a network that has random weights on its connections between nodes.

2. Supervised Learning:

You have two groups of siblings, and you want the network to learn which person belongs to which group.

3. For Each Person:

Take one person at a time.

Input information about that person into the network.

4. Compare Output:

Compare the network’s output (what it thinks about the group the person belongs to) to the correct or required output (the actual group the person belongs to).

5. Calculate Error:

Determine the difference or error between the network’s output and the correct output.

6. Adjust Weights Backwards:

Starting from the output layer and moving backward through the layers:

Modify the connection weights to reduce the error.

Focus more on connections that contributed more to the error.

7. Repeat for Each Person:

Go through this process for each person in your dataset.

8. Iterations:

Repeat this process for multiple rounds or iterations, going through the entire dataset each time.

9. Network Improvement:

With each iteration, the network adjusts its weights to get better at assigning the correct group to each person.

10. Training Completion:

Continue this process until the network reaches a point where it consistently assigns the correct groups based on the input.

Question 5

Does adding layers in an Artificial Neural Network improve performance?
- Answer
- Analogy with neurons

Answer 5

Answer:
No, this depends on the task.
* Deep networks allow for hierarchical
representation.
* Together with convolution, one of the causes of the
successes of deep networks

Analogy with neurons:
* There is a minimum number of neurons needed to
approximate any specific function, but simply
adding more neurons does not work well.
* Instead, specific changes in topology are needed,
e.g. deep networks.