Neural Network

[Aziz]

Is there any reason why you don't post it here?

Yes, there is a reason.
How about fairness I asked recently?

Aziz

[Qiaozhi]

Yes, there is a reason.
How about fairness I asked recently?

Aziz

I suppose that's your prerogative.

Neural networks (NN) could theoretically be used in metal detectors, but whether the added complication would be a step-function increase in performance is open for debate. Dave Johnson obviously considers the idea to be of little practical value, and I tend to agree. I have some small experience using back-propagation neural nets, and I can see that the concept is do-able, but only of academic interest. Neural networks are very useful for finding patterns in imprecise data, which you could argue is the case for buried metal targets, as there is some correlation between the signals received by the metal detector and the actual target, and a trained network can also make decisions on data it has never seen before by using the knowledge gained during training.

However, you do have to be careful when training a neural network. For example, there is a story concerning the U.S. Military, where they attempted to use an NN to identify tanks in the battlefield as either friend or foe. To do this they input photos of American and Russian tanks, and used these to train the network. It was 100% successful at identifying the targets when tested on the training data, but in practice it was completely useless when used on any photos outside of the training set. After some analysis, it was noticed that all the American tanks had been photographed on a sunny day, and the Russian tanks had been photographed on a cloudy day. The NN was simply identifying the different weather conditions.

Neural nets can also provide an indication of whether there is any correlation to be found within a dataset. A good example is to take a database of recent Lottery numbers and use these as training data. The first thing you discover is that the network is unable to converge, especially if the dataset is quite large. The hypersurface does not contain a true global minimum, and even if you stop the training process once the network shows no further improvement in the network error, the predicted result for next week's Lottery will not make you a rich man. This is because the Lottery results bear no correlation to the previous results, despite what many so-called Lottery prediction programs would have you believe.

So, no doubt you would find some correlation in the metal detector data, but is it worth the effort?

[Aziz]

I suppose that's your prerogative.

Neural networks (NN) could theoretically be used in metal detectors, but whether the added complication would be a step-function increase in performance is open for debate. Dave Johnson obviously considers the idea to be of little practical value, and I tend to agree. I have some small experience using back-propagation neural nets, and I can see that the concept is do-able, but only of academic interest. Neural networks are very useful for finding patterns in imprecise data, which you could argue is the case for buried metal targets, as there is some correlation between the signals received by the metal detector and the actual target, and a trained network can also make decisions on data it has never seen before by using the knowledge gained during training.

However, you do have to be careful when training a neural network. For example, there is a story concerning the U.S. Military, where they attempted to use an NN to identify tanks in the battlefield as either friend or foe. To do this they input photos of American and Russian tanks, and used these to train the network. It was 100% successful at identifying the targets when tested on the training data, but in practice it was completely useless when used on any photos outside of the training set. After some analysis, it was noticed that all the American tanks had been photographed on a sunny day, and the Russian tanks had been photographed on a cloudy day. The NN was simply identifying the different weather conditions.

Neural nets can also provide an indication of whether there is any correlation to be found within a dataset. A good example is to take a database of recent Lottery numbers and use these as training data. The first thing you discover is that the network is unable to converge, especially if the dataset is quite large. The hypersurface does not contain a true global minimum, and even if you stop the training process once the network shows no further improvement in the network error, the predicted result for next week's Lottery will not make you a rich man. This is because the Lottery results bear no correlation to the previous results, despite what many so-called Lottery prediction programs would have you believe.

So, no doubt you would find some correlation in the metal detector data, but is it worth the effort?

Hi Qiaozhi,

nice weather detection example. Brilliant. The NN did work. I like it.
This is a good example of one of the pitfalls in NNs.

Is it worth at all?
Metal detection: No, metal detection is trivial and a well-known science. You know, how to process the data.

But there are many many non-trivial problems that might be challenged. If you don't know, how to solve a problem. If you don't know, what parameters do affect the problem. If you don't know, whether a problem can be solved or not. In this case, the NN is a very very useful data mining tool.

Nevertheless, it causes a lot of work. And it isn't justifying the lots of effort in many cases.

Cheers,
Aziz

[moodz]

Hi Qiaozhi,

nice weather detection example. Brilliant. The NN did work. I like it.
This is a good example of one of the pitfalls in NNs.

Is it worth at all?
Metal detection: No, metal detection is trivial and a well-known science. You know, how to process the data.

But there are many many non-trivial problems that might be challenged. If you don't know, how to solve a problem. If you don't know, what parameters do affect the problem. If you don't know, whether a problem can be solved or not. In this case, the NN is a very very useful data mining tool.

Nevertheless, it causes a lot of work. And it isn't justifying the lots of effort in many cases.

Cheers,
Aziz

don't discount neural nets so easily ....after all what is that came up with all metal detector technologies ... A neural net (your brain LOL ). See link for clues on neural net DSP .....
http://www.asel.udel.edu/icslp/cdrom/vol1/289/a289.pdf