The following MPEG movies were put together for a presentation which I did at ECAL'97 for the paper:
Activity-Based Pruning in Developmental Artificial Neural
Networks
Rust A G, Adams R, George S and Bolouri H.
In Husbands P and Harvey I (eds),
Proceedings of the 4th European Conference on Artificial Life
(ECAL'97), Cambridge, MA, MIT Press, 224-233, 1997
The movies aim to visually demonstrate some of the key principles of the developmental simulator. The second movie records the development of the network shown in Figure 4(a) of the paper and the third movie the two neurons in Figure 6.
To play the movies I use mpeg_play running under Linux.
This movie simply illustrates how neurons branch and interact with each other. Gradients of chemical gradients (invisible in the movies) cause growing axons and dendrites to navigate towards each other. Where axons and dendrites make contacts synapses are formed; these are represented by the red diamonds. Axons and dendrites on the periphery of the chemical gradients are too distant from their intended attractive gradients and as a result explore their local environments.
This movie is simply a more complex version of the first movie. The movie demonstrates the extent and complexity of potential growth in the developmental model.
In this example many of the lower layer neurons develop neural trees which do not form connections with the single, upper layer neuron. This results in overgrowth in the model, a common phenomenon in biological development. Based on the cummulative effects of activity-based events the structure and connectivity of this network are subsequently pruned - see the paper for further details.
The self-regulation of neural growth is demonstrated in this movie. The two neurons have been extracted from the network in movie 2 above. Both neurons are growing towards the single upper layer neuron, which is unshown in the movie but would be positioned in the top righthand corner of the frame.
The righthand neuron forms synapses with the upper neuron after a number of time steps. Due to activity-based feedback mechanisms via the formed synapses, the growth of the neuron is gradually reduced until the neuron stops growing completely. The feedback mechanisms also incrementally reinforce the strengths/weights of the formed synapses. This is visually represented by the enlargement of the synapses in the movie.
The lefthand neuron does not make connections with the upper neuron and it's growth is thus unregulated. For the duration of the movie it's growth remains unchecked such that it develops a far larger neural tree than the righthand neuron.
The movies were generated using VRML (the Virtual Modelling Reality Language) to visualise the networks and a nice little package called xmpegrab from COGS, Sussex University.