CODEWITHSUNDEEP
testingneural-networkartificial-intelligencestochastic-process
user7190127
user7190127

Reputation:

How to check Artificial Neural Network results not due to chance

I fully understand the theory behind an ANN (in this case, feed forward with back propagation). As the network learns, the weights are adjusted accordingly in order to give the correct results. However, due to the stochastic element involved i.e using random weights to initialise the network, how can I check that the the results produced are not just due to chance / pure coincidence?

Upvotes: 0

Views: 98

Answers (1)

Mark Adelsberger
Mark Adelsberger

Reputation: 45689

It seems like a rather philosophical question, doesn't it? For a non-trivial case the specific weight values will be driven in part by chance... and so what? As you know, a characteristic of an ANN is that the "reasoning process" cannot be traced; either it gets the correct answer, or it doesn't. Through training you refine a function that you hope will get the right answer. If that function produces correct results reliably, then it makes no difference that chance was an element in the derivation of the function.

It is possible that an ANN would get the correct result on every problem in the training set and still get wild results on any problem it hasn't seen before. This could mean it was overparameterized in the first place, so that it only "memorized" the training set. It could mean that the training set was insufficient to represent the range of the problem. In each of those cases, chance will determine what partially-correct function you've ended up with, and what kind of wrong results you get for additional problems.

But if you have an ANN well-suited to a problem and a training set sufficient to prime it, then it will produce correct results reliably. Chance played a role in deciding exactly what function it embodies, but the selected function - which is opaque anyway - works, so who cares?

Upvotes: 1

Related Questions