Microscopia de fuerza atómica (AFM)

The basic “architecture” of the neural network refers to its arrangement of

arrangement of the neural network as follows: neuron, input layer, hidden layer(s), output layer, and connection weights. A schematic architecture of ANN and its four parts is shown in Figure 3.1. A brief description of each part is given in this section.

3.2.1 Neuron

The basic building block of the network system is the neuron, the cell that communicates information to and from the various parts of the body (Figure 3.2 shows a neuron with its

different constituents). Abrahart & See (2000) stated that that the biological neuron consists of three main parts, namely:

• A cell body called the soma

• Several spine-like extensions of the cell body called dendrites

• A single nerve fiber called the axon that branches out from the soma and connects to many other neurons

The axons and dendrites are considered to be responsible for transmitting signals to the neuron. Figure 3.3 represents an artificial neuron in its simplest form. The incoming lines in Figure 3.3 represent dendrites. Each line carries a signal from another neuron. The body

represents the soma and the output represents the axon, which in its turn branches to interconnect with other neurons (Ham & Kostanic, 2000). All artificial neurons interconnect to each other form what is called an ANN. The McCulloch-Pitts neuron (Figure 3.4) is the most commonly used neuron model.

According to Ham & Kostanic (2001), each artificial neuron forms a node in the larger neural network and is constructed of the following basic elements:

• Synapses or connection links send input from one node to another in an ANN. Each synapse has its own weight or strength. A positive weight indicates an excitatory

synapse; a negative weight indicates an inhibitory one.

• An adder or linear combiner sums the weighted input signals from other nodes transmitted via the synaptic connections.

• The activation function limits the amplitude of the output of the artificial neuron. The activation functions, which are described in greater detail in section 3.6, can be continuous- values, binary (with range [0,1]) or bipolar (with range [-1,1])

• A bias (θi) may also be present. The bias increases or decreases the net input of the

activation function.

The neuron has one or more inputs and produces one output. The inputs simulate the stimuli/signals that a neuron gets, while the output simulates the response/signal which the neuron generates. The output is calculated by multiplying each input by a different number (called weight), adding them all together, then scaling the total to a number between 0 and 1.

3.2.2 Input Layer

Anderson & McNeil (1992) indicated that the input layer is the least complex of all the layers because no mathematical calculations occur at this level. Before beginning, the number of inputs and relevance of the inputs must be decided. Inputs that are believed to have no relevance on the output should be eliminated. Thus, available input data that affect the output are fed to the network. The performance of the network depends on the number of inputs. The input layer receives and processes information and forwards it to the hidden layer.

3.2.3 Hidden Layer(s)

It is at this layer that all the calculations occur. The numbers of hidden layers vary but there is always at least one hidden layer in every network. Each layer is composed of a set of neurons. Each layer is interconnected in such a way that the first layer passes information to the second layer, the second layer to the third, and so forth (Huang & Dong, 1992). This is done via

connection weights (Figure 3.5). Connection weights connect each neuron in a certain layer to every single neuron in the next layer. The value of that weight is responsible for adjusting the output value of the neuron.

Each processing element in a specific layer is fully or partially connected to many other processing elements via weighted connections. The scalar weights determine the strength of the connections between interconnected neurons. A zero weight refers to no connection between two neurons and a negative weight refers to a prohibitive relationship. From many other processing elements, an individual processing element receives its weighted inputs, which are summed, and a bias unit or threshold is added or subtracted. The bias unit is used to scale the input to a useful range to improve the convergence properties of the neural network. The result of this combined summation is passed through a transfer function (e.g. logistic sigmoid or hyperbolic tangent) to produce the output of the processing element.

3.2.4 Output Layer

Najjar & Basheer (1996) indicated that the output layer in a network is a layer containing one or more output neurons. An output neuron will compute a value for a certain parameter or variable.

Input value to node k: k jk j j

I =

∑

The output values Oj that leave a node j on each of its outgoing links are multiplied by a

weight, wj. The input Ik to each node k in each middle and output layer is the sum of each of its

3.2.5 Connection Weights

Connection weights are the interconnecting links between the neurons in the layers constituting the network. Each neuron in a certain layer should be connected to every single neuron in the next layer by a connection weight. The value of that weight is responsible for adjusting the output value of the neuron. The magnitude of the weighted connection is directly related to the strength of that connection (Romaniuk, 1995). Signals travel between neurons over weighted connection links. The weight assigned to the connection is multiplied to the signal that is transmitted. Each connection link has an associated weight, which, in a typical neural network, is multiplied to the signal that is transmitted. The process of training a neural network involves the adjustment of the weights based on the given learning rule (Ham & Kostanic, 2000). The overall net input consists of the sum of the weighted connections (product of the weight times the signal).