Biological analogy
Artificial Neural Networks (ANN) form a
family of inductive techniques that mimic
biological information processing. Most animals
possess a neural system that processes
information. Biological neural systems consist of
neurones, which form a network through synaptic
connections between axons and dendrites. With
these connections, neurones exchange
chemo-electrical signals to establish behaviour.
Neural can adapt their behaviour by changing the
strength of these connections; more complex
learning behaviour can occur by forming new
connections. Changes in connections, and new
connections, grow under influence of the signal
exchange between neurones.
|
[click to enlarge]
|
|
|
Figure 7.
Biological neural tissue.
Neural cells (a), axons (b) and dendrites
(c) are visible. Where axons touch
dendrites or nuclei, synaptic contacts
are formed.
|
Figure 8.
Layered artificial neural network.
On the left, a pattern is coded.
The activation rule propagates the
pattern to the right side through the
central layer, using the weighted
connections. From here, the output can be
read. Based on the input-output relation,
the learning rule may update the weights
|
Artificial implementation
Artificial neural networks are simplified models of
the biological counterpart. They consist of the
following elements:
- Processing
units (models of neurones)
- Weighted
interconnections (models of neural
connections)
- An activation
rule, to propagate signals through the
network (a model of the signal exchange
in biological neurones)
- A learning
rule (optional), specifying how weights
are adjusted on the basis of the
established behaviour
The most common
ANN are structured in layers (see Figure
8), with an input layer where data is
coded as a numeral pattern, one or more hidden
layers to store intermediate results, and an
output layer that contains the output result of
the network.
Neural Network
Model Representation
In contrast to
symbolic techniques such as decision trees
(see link above) the representation of
knowledge in a neural network is not easy to
comprehend. The set of weights in a network,
combined with input and output coding, realises
the functional behaviour of the network. As
illustrated in Figure 3, such data does not give
humans an understanding of the learned model.
Some alternative
approaches for representing a neural network are:
- Numeric:
show the weight matrix; not very
informative, but illustrates why a neural
network is called sub-symbolic: the
knowledge is not represented explicitly,
but contained in a distributed
representation over many weight factors.
- Performance:
show that a neural network performs good
on the concepts in the data set, even
better than other techniques, and do not
try to interpret de model
- Extract
knowledge: techniques exist to
extract the model captured in a neural
network, e.g. by running a symbolic
technique on the same data or by
transferring a NN model into an
understandable form. Techniques exist to
extract statistical models and symbolic
models from NN.
|
