Scribd
Upload
22 views

M. Raihan

The document discusses neural networks and how perceptrons can be trained. It explains how perceptrons work with inputs, weights, biases and thresholds. The learning rule for updating weights is also explained. Limitations of single layer perceptrons are that they can only learn linearly separable problems.

Uploaded by

Sabuj Dhali
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
22 views

M. Raihan

The document discusses neural networks and how perceptrons can be trained. It explains how perceptrons work with inputs, weights, biases and thresholds. The learning rule for updating weights is also explained. Limitations of single layer perceptrons are that they can only learn linearly separable problems.

Uploaded by

Sabuj Dhali
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 20

M.

Raihan
Email: [email protected]
Neural Networks
Neural Networks
Can put lots of McCulloch & Pitts neurons
together.
Connect them up in any way we like.
In fact, assemblies of the neurons are capable of
universal computation.
Can perform any computation that a normal computer can.
Just have to solve for all the weights wij

21-Sep-17 3
Neural Networks

21-Sep-17 4
The Perceptron Network

21-Sep-17 5
Training Neurons
Adapting the weights is learning
How does the network know it is right?
How do we adapt the weights to make the network
right more often?
Training set with target outputs.
Learning rule.

21-Sep-17 6
A Simple Perceptron
It’s a single-unit network
Change the weight by an
amount proportional to
the difference between
the desired output and
the actual output.
21-Sep-17 7
Updating the Weights
 Wij Wij+ Wij
We want to change the values of the weights
Aim: minimize the error at the output
If E = t-y, want E to be 0
Use:

21-Sep-17 8
The Learning Rate
ɳ controls how much to change by weights.
If we could miss it, ɳ = 1.
Weight change a lot, whenever the answer is wrong.
Makes the network unstable.

Small ɳ
Weights need to see the inputs more often before they change
significantly.
Network takes longer to learn.
But, more stable network.
21-Sep-17 9
Bias Input
 What happens when all the inputs to a neuron is zero?
It doesn’t matter what the weights are
The only way that we can control whether neuron fires or not is
through the threshold.
That’s why threshold should be adjustable.
Changing the threshold requires an extra parameter that we need to
write code for.
We can add an extra input weight to the neuron, with
the values of the input to that weight always being fixed.
h = 𝒎 𝒊=𝟏 𝒙𝒊 𝒘𝒊 21-Sep-17 10
Biases Replace Thresholds

21-Sep-17 11
Training a Perceptron

21-Sep-17 12
Training a Perceptron

21-Sep-17 13
Training a Perceptron

21-Sep-17 14
Linear Separability

21-Sep-17 15
More Than One Input Neuron
The weights for each neuron separately describe a straight line.

21-Sep-17 16
Perceptron Limitations
A single layer perceptron can only learn
linearly separable problems.
Boolean AND function is linearly separable, whereas
Boolean XOR function (and the parity problem in
general) is not.

21-Sep-17 17
Linear Separability

21-Sep-17 18
What Can Perceptrons Represent?

21-Sep-17 19
Thank You

21-Sep-17 20

You might also like