Scribd
Upload
37 views

6 CS1AS16 Neural Computing

Neural networks are modeled after the human brain and consist of interconnected nodes called neurons that can learn from examples. A neural network uses a learning algorithm to adjust the connections between neurons based on training data to perform tasks like classification, prediction, and control. Common neural network architectures include the perceptron, feedforward networks with multiple hidden layers, and recurrent neural networks with feedback connections.

Uploaded by

Anna Abcxyz
Copyright
© © All Rights Reserved
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
37 views

6 CS1AS16 Neural Computing

Neural networks are modeled after the human brain and consist of interconnected nodes called neurons that can learn from examples. A neural network uses a learning algorithm to adjust the connections between neurons based on training data to perform tasks like classification, prediction, and control. Common neural network architectures include the perceptron, feedforward networks with multiple hidden layers, and recurrent neural networks with feedback connections.

Uploaded by

Anna Abcxyz
Copyright
© © All Rights Reserved
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
You are on page 1/ 3

CS1AC16 NEURAL COMPUTING

A traditional computer has a big fast processor that can perform millions operations per second.
The human brain has a lot of small and slow processors – neurons. What makes the brain powerful is
that is has millions of neurons that are connected together and can learn.
Neural computing tries to mimic the model of a brain.

“Neural computing is the study of networks of adaptable nodes which, through a process of learning
from task examples, store experiential knowledge and make it available for use.”

Neural network schematic:

Neurons make calculations basing on the input values.

As neural networks are based on the functioning of a human brain, it is useful to know the basics of
how a brain works:
 A neuron receives input from other neurons via dendrites
 If it's sufficiently stimulated, it generates an impulse
 The impulse travels via the axon to other neurons
 Junctions between the neurons are called synapses
The structure of a neuron:

When an impulse reaches a synapse, it releases


neurotransmitters that stimulate dendrites of the
receiving neurons. The connection strength (weight)
of the synapse affects how the neurons are
stimulated. They can be:
 Excitatory (activity aids firing)
 Inhibitory (activity suppresses firing
The neuron sums all the signals it receives and fires
an impulse if this sum exceeds a threshold.
The connection strength change as people learn.
Human brain is very complex: it has 10 12 neurons and 1015 synapses (there is 100 times more neurons in
a brain than trees in Amazon forest).

Artificial neural network(ANN) is a model of what goes one in the brain but a simplified one:
 A model of neuron – a simplified version of a real one
 These neurons can be connected in many different ways
There can be many types of ANN.
What can they do?
o Classification – for a given input say it belongs to class A or B
o Association – map the input to an output
o Prediction – calculate output for a given input
o Control – generate control signal

The first model of a neuron was the McCulloch Pitts neuron. Many types of neural networks evolved
from it.

Perceptron

It has an input layer of values (light) and a layer of


neurons (dark). It is sometimes called single layer
perceptron network. The green neurons multiply the
inputs by connection strengths and fire if the total is
bigger than the threshold.

Minsky and Papert showed that a single layer is inadequate – we can have multiple layer network:

From the input layer (on the left) the


inputs go to the hidden layer (the
middle one), from where they go to the
output layer (on the right).
A network can have multiple hidden
layers.

Feedforward network – data goes forwards until it comes out – like in the examples above.

A network must be set up properly in order to do something useful – meaning the weights between the
connections must be set.
The network can learn:
o Supervised learning - through training data – you provide inputs, the network
calculates outputs but the expected outputs are also known – the network can adjust
the weights based on the error:
o Unsupervised learning – the data is provided to all neurons and they work out by
themselves how to operate – groups of them start recognizing similar values (self
organizing nets). The neurons are often arranged in a rectangular grid and neurons in
one area respond to certain inputs.
 Example: network trained to recognize colors:
Here, similar colors are close to each other.

Recurrent neural networks(RNN) – networks with feedback. They're more complicated than
feedforward networks.

In this example the output is calculated using


input values and the previous output – they
have memory.

Feedforward networks calculate output based


on current inputs. Sometimes the output is
affected by the previous inputs – for example,
word prediction. RNN can be used in this case.

You might also like