Downloads

Both software and movies are available for downloading. They are all developed using a Microsoft operating system, but some of them are useable when transferred to other operating systems.

Software

• Version II_III of the perspex machine is implemented in Pop11. It can be run on both Microsoft and Unix operating systems. This version of the machine was used to develop the examples in the papers Perspex Machine II and Perspex Machine III. It does not have documentation. It is not a generic implementation of the perspex machine. It will be of use only to skilled programmers.

• Isabelle/HOL proof files supporting the paper Perspex Machine VIII. This code will be of use only to mathematicians familiar with Isabelle/HOL.

Movies

• FibonacciPerspex shows a perspex neural net compiled from a C program that calculates the first ten Fibonacci numbers.
  
  void main() {double i, j, k, l; i = 0; j = 1; k = 1; l = 0; for (l = 0; l < 10; l = l + 1) {k = i + j; i = j; j = k;} k = k;}

• The first frame of the movie shows the initial condition of the neural net. The neurons are shown in the standard perspex colours: read synapses red, body orange, write synapse green, and jump synapses blue. Processing starts in the bottom left of the net. The neuron at the top left controls the number of Fibonacci numbers that are computed. Moving this neuron higher up causes the net to compute more Fibonacci numbers, lowering it causes the net to compute fewer Fibonacci numbers. The net starts by growing some local variables vertically upwards on the left hand side. It then grows each Fibonacci number rightward, storing the result in one of the local variables. When the net halts this variable holds the tenth Fibonacci number.
  
  In principle, any C program can be compiled into a perspex neural net. The program then inherits the properties of perspex neural nets but, unless these are accessed, it executes exactly what the C source program instructed.
  
  If the neural net were modified in some way, say, by applying filtering or a genetic algorithm, the result would be a new program related to the computation of Fibonacci numbers. If the modification were extreme the relationship would be tenuous.

• DijkstraPerspex shows a net with over 600 perspex neurons that implements Dijkstra's solution to the Travelling Salesman problem. This net was compiled from the C source code for Dijkstra's algorithm.
  
  The movie shows the entire net with all but the recently used neurons fading out over time. When the net halts it has correctly computed the shortest distance between the cities visited by the travelling salesman.
  
  The vertical part is mostly local variables. The cubical block of neurons to the left is the arrays used to hold data about the position of cities. The fibres to the right are mostly the evaluation function. The answer is recorded in one of the local variables.