Exam Question Notes

Question 1

Will the increase of the number of hidden nodes in a multilayer perceptron improve the
generalisation capability of the model? Why (not)?

Answer 1

Increased Capacity: More hidden nodes allow the network to learn more complex patterns and relationships in the data.
Underfitting Fix: If the model is too simple (underfitting), adding nodes can help it better capture the underlying structure.

Overfitting Risk: Too many hidden nodes can cause the model to memorize the training data rather than generalize to unseen data.

Question 2

How can you determine the architecture of a NN for a given data set?

Answer 2

1) Inputs = predictors
2) Outputs = enough to rep. classes in binary
3) Hidden layers = number of regions of data
4) hidden nodes = data boarders (eg 4 for a square)

One got one example to go off here

Question 3

What methods can be employed to prevent overfitting when iterativley training a model?

Answer 3

1) Early stop
2) Collect more training data
3) Better validation

Question 3

By using sensory data from multiple sources could build a better diagnosis model by fusing
the data. What are the three typical strategies in sensor fusion and how to implement?

Answer 3

To achieve sensor fusion, the three typical fusion strategies include data level, feature level, and decision level.
* For data level fusion, the raw data from the sensors without feature extraction would be input to a model that could deal with the data from different sources, possibly with different dimensionality.
* For feature level fusion, features are extracted first from each data source. Then the features are combined as input to the developed machine learning model.
* For decision level fusion, multiple classification models are built first that each works on a single source of data. Decisions from each model are then fused by some mechanism (such as majority voting or weighted average of the classification probabilities) to output the final classification result.