17.01.2013 Views

Chapter 2. Prehension

Chapter 2. Prehension

Chapter 2. Prehension

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

384 A pp e n dic e s<br />

C.2 Artificial Neural Networks<br />

Artificial neural networks consist of processing units. Each unit<br />

has an activation state and are connected together through synapses’ in<br />

some pattern of connectivity. Rules govern how information is<br />

propated and how learning occurs in the network. The processing<br />

units can represent individual neurons or else concepts (that can either<br />

be individual cells themselves or groups of cells). The issue, more<br />

importantly, for these models is how a computation is performed<br />

without regard so much as to at level it is working. In the next<br />

subsections, we describe some of the decisions that a network<br />

designer must make, and what are the reasons and advantages for<br />

doing so.<br />

C.<strong>2.</strong>1 Activation functions and network topologies<br />

Different models are used within neural network models for the<br />

processing units. The most common is the McCulloch-PittS type<br />

neuron, where each unit simultaneously takes a weighted sum of its<br />

inputs (dendrites) at each network time step At, and if threshold is<br />

reached, the cell body discharges at a certain firing frequency. The<br />

output of such a neuron is seen in the left side of Figure C.l. If we<br />

consider the weighted sum of a neuron’s incoming signals to be its<br />

input, I, and its state to be u, then a McCulloch-Pitts neuron operates<br />

according to the assignment u = I. The result is then thresholded to<br />

produce its output. An alternative to this discrete model is the leakv<br />

interrrator model, where the neuron’s behavior is described by a frrst-<br />

order linear differential equation:<br />

where z is the unit’s time constant. In simulating such a neuron, z<br />

must be greater than the simulation time step At. As the time constant<br />

approaches the simulation time step, this rule approaches the<br />

McCulloch-Pitts rule. The output of such a neuron is seen in the right<br />

side of Figure C.l. If z = dt, we see that du = -u + I. So, in the<br />

simulation,<br />

‘A synapse is a small separation between the axon (output fiber) of one neuron and the<br />

cell body or fibers of another neuron. A neuron can synapse onto itself as well.

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!