rOmV4 - Control Systems Neural Networks

To Train a Network

This is one simple method.

Make some random changes to the strengths of the links between the processors.

Test the new network.

If it performs better, keep the changes.

If it is the same or worse, discard the changes and try again.

Repeat these steps until the network performance is good enough.

A method is needed to work out a performance score for measuring the quality.

Comparing Conventional and Neural Network Computation

Inputs

Processor

Programming

Memory

Applications

Solves well defined problems like

payroll or
bank account management.
Can do precise maths.
Good for solving problems with small numbers of inputs and where the algorithm (set of rules) for solving the problem is well understood.

Better at predictive problems such as ...

Weather forecasting
Image recognition. Solves problems like image recognition (spot the criminal or terrorist in the shopping mall).
Good for solving problems with vast amounts of input data where the rules for reaching the answer are fuzzy or not well defined.

Subject Name Level Topic Name Question Heading First Name Last Name Class ID User ID

Q: qNum of last_q Q ID: Question ID Score: num correct/num attempts Date Done

Question Text

image url

Help Link

Add Delete Clone Edit Hardness

Help Text

Debug

You can attempt a question as many times as you like.
If you are logged in, your first attempt, each day, is logged.
To improve your scores, come back on future days, log in and re-do the questions that caused you problems.
If you are logged in, your most recent wrong answers get remembered. This might help you and your teacher to correct your understanding.
In the grade book, you can delete your answers for a topic before re-doing the questions. Avoid deleting unless you intend re-doing the questions very soon.

Hosted at linode.com - London

Please report website problems to Neil

	Conventional Computer	Neural Network
Inputs	A few inputs like a keyboard, mouse or bar code scanner.	Many inputs, possibly thousands. For example an image recognition system would have inputs for each pixel.
Processor	A single processor which deals with all the data. Some computers have two or four processors but rarely more.	Many processors, possibly thousands. Each processor is much simpler than a conventional computer processor. Each processor deals with a small part of the data. The processors work in parallel.
Programming	Must be programmed by a programmer.	Must be trained. Neural networks learn instead of being programmed. Learning takes place by links being made stronger or weaker.
Memory	Data stored centrally (ROM, RAM, hard disk)	Data stored as weightings in the network.
Applications	Solves well defined problems like payroll or bank account management. Can do precise maths. Good for solving problems with small numbers of inputs and where the algorithm (set of rules) for solving the problem is well understood.	Better at predictive problems such as ... Weather forecasting Image recognition. Solves problems like image recognition (spot the criminal or terrorist in the shopping mall). Good for solving problems with vast amounts of input data where the rules for reaching the answer are fuzzy or not well defined.