Would an optimum combination of weights for a given topology necessarily be the the optimum for a different topology ?

  • 1
    $\begingroup$ What are your thoughts? Why would you expect weights to carry over to other nets? $\endgroup$ – Raphael Oct 22 '16 at 8:20
  • 2
    $\begingroup$ I don't understand your question. The weights are applied to edges of a graph. What does it even mean to put "the same weights" on a different graph? $\endgroup$ – David Richerby Oct 22 '16 at 9:27

No, an optimum combination of weights for a given topology not necessarily be the the optimum for a different topology.

Weight of a given neural network depends on the structure of the network and training data which vary network to network (or problem to problem).

Always the weights of a network are adjusted to match our desire output to achieve our goal.

Network topology also have a critical role in selection an optimum combination of weight.

Network with one hidden layer:

Network with one hidden layer

Network withe two hidden layer:

Network withe two hidden layer

In above figures shows classification of a linearly separable data on two different network topology.

Here, thick dash line shows strong connection(higher weight) and thin dash line shows poor connection(lower weight).

You can visualize how weights get changes when we fed same data to two different network.

I recommend you to go to http://playground.tensorflow.org/ and play around different setting and observe the changes.

  • $\begingroup$ Alwyn, if you're going to edit LaTeX into posts, could you at least turn all the maths into LaTeX? In this edit, you left us with the choice of either rejecting (which I did), accepting to give a post with inconsistent formatting, or spending time finishing the edit that you should have done. This just wastes our time. Also, please don't reformat off-topic posts. If the post is off-topic, no amount of reformatting will make it on-topic. Again, it's just wasting our time. $\endgroup$ – David Richerby Oct 28 '16 at 11:21

Not the answer you're looking for? Browse other questions tagged or ask your own question.