Popular recipes tagged "neural_networks"http://code.activestate.com/recipes/tags/neural_networks/popular/2012-10-02T16:18:36-07:00ActiveState Code RecipesArtificial Neuroglial Network (ANGN) (Python) 2012-10-02T16:18:36-07:00David Adlerhttp://code.activestate.com/recipes/users/4182015/http://code.activestate.com/recipes/578242-artificial-neuroglial-network-angn/ <p style="color: grey"> Python recipe 578242 by <a href="/recipes/users/4182015/">David Adler</a> (<a href="/recipes/tags/artificial_intelligence/">artificial_intelligence</a>, <a href="/recipes/tags/genetic_algorithm/">genetic_algorithm</a>, <a href="/recipes/tags/genetic_algorithms/">genetic_algorithms</a>, <a href="/recipes/tags/neural/">neural</a>, <a href="/recipes/tags/neural_networks/">neural_networks</a>). Revision 5. </p> <p>This is an attempt at emulating the algorithm from these scientific articles:</p> <ol> <li><a href="http://www.hindawi.com/journals/cmmm/2012/476324/">2011 - Artificial Astrocytes Improve Neural Network Performance</a></li> <li><a href="http://www.plosone.org/article/info%3Adoi%2F10.1371%2Fjournal.pone.0019109">2012 - Computational Models of Neuron-Astrocyte Interactions Lead to Improved Efficacy in the Performance of Neural Networks</a></li> </ol> <p>The objective of the program is to train a neural network to classify the four inputs (the dimensions of a flower) into one of three categories (three species of flower), (taken from the <a href="http://archive.ics.uci.edu/ml/datasets/Iris">Iris Data Set</a> from the UCI Machine Learning Repository). This program has two learning phases: the first is a genetic algorithm (supervised), the second is a neuroglial algorithm (unsupervised). This ANGN is a development of a previous program only consisting of a genetic algorithm which can be found <a href="http://code.activestate.com/recipes/578241-genetic-algorithm-neural-network-in-python-source-/">here</a>.</p> <p>The second phase aims to emulate astrocytic interaction with neurons in the brain. The algorithm is based on two axioms: a) astrocytes are activated by persistent neuronal activity b) astrocytic effects occur over a longer time-scale than neurons. Each neuron has an associated astrocyte which counts the number of times its associated neuron fires (+1 for active -1 for inactive). If the counter reaches its threshold (defined as <code>Athresh</code>) the astrocyte is activated and for the next x iterations (defined as <code>Adur</code>) the astrocyte modifies the incoming weights to that particular neuron. If the counter reached a maximum due to persistent firing the incoming weights are increase by 25% for the proceeding <code>Adur</code> iterations; conversely if the counter reached a minimum due to persistent lack of firing the weights are decreased by 50% for the following <code>Adur</code> iterations). For a detailed description of the algorithm see the linked articles. For a general understanding of how this program was coded look at the pseudo-code/schematic <a href="http://commons.wikimedia.org/wiki/File:ANGN_schematic.png">here</a>.</p> <p>Any comments for improvements are welcome. There are several issues in this program which require addressing, please scroll down below code to read about these issues.</p>