Sample Questions from Week 4 - SOM

1. Network Design and training
1. Draw a SOM network that could be used to model a simple syllable map (ba, da, ga, bi, di, gi, etc).
   Assume that the inputs to your network can be represented using units for 6 vowels and 6 consonants.
   Let the map layer have 100 units, organised in a 2D sheet.
2. How many layers does your SOM network have?
3. How many units are needed in each layer?
4. How many weights are in your network? Explain your workings.
5. What learning rule would be appropriate?
6. What would be appropriate initial values for the weights?
7. What sort of data would your network be modelled on?
   
   
2. Properties of SOMs
1. Self-organizing maps (SOMs) have been used to study the organization of orientation detectors in the visual cortex. Explain how a SOM network could learn to represent oriented lines from a simple retina of 3x3 units.
2. What effect does changing the neighbourhood size have on performance of a SOM?
3. Can a SOM represent patterns that it has not seen during training? Justify your answer.