Connectionist Modelling Exercise 2
This exercise is to be your own effort, without the input or collaboration of your fellow students. It is open book, i.e. you may use books, notes or the web in doing it, but you may not confer or consult with any other person.
The exercise is due on Friday, march 31st, 2017. You should hand up a written (printed) answer to all questions below. Please also submit an electronic copy for my records.
Before attempting this exercise, you should have completed the first six labs...
Recognizing LED digits
In this mini project you have to implement the task of recognising the LED digits 0 - 9 in a connectionist network of your design.
Your task is straight forward: Build a connectionist model that is able to recognise all the possible 10 digits from an LED display. You can use whichever set-up you like, as long as it works. You will have To decide what your input patterns are, and what your output patterns are. A straightforward choice would have 7 input nodes (A-G, see picture below) and 10 output nodes (numbers 0-9, localist coding), but you are welcome to find any other way of representing inputs and outputs. An input node is on (value of 1) when the corresponding diode is lit, off (value of 0) otherwise. For the digit `2', for example, all inputs except A and G should be on.
Here is food for thought: what constitutes a test of such a network, beyond the 10 patters on which it is trained? There is no obvious or simple answer, and so you may decide to test it in various ways. Can you design a test that resembles a real world situation, e.g. low voltage so that an element is neither clearly on nor off, or perhaps a broken LED. Design of a test scenario is up to you, but you should try to relate your test to the physical situation being modelled.
What to deliver
Your task is to develop and document a simulation that solves the above problem. You will have to make decisions about what a suitable training set, and a suitable test set, should be. Your assignment should be explicit about any choices you make about representations, set make up, etc.
Your pattern files will be plain text files (no MS Word documents!) and the pattern files you have used up to now will serve you as a model.
You may choose to seek a network that is maximally efficient in terms of training, but you don't have to. You do need to document all relevant parameter values, and the performance of your trained network.
Finally, you should include some (well-written) text that conveys the challenges and issues you faced in completing this exercise.