-1

While I am trying to get the deep learning flow, I can not find out one detail -> when I reach an error on every neuron (in backpropagations flow), what is next I should do with that all errors. The calibration of model is about adjusting the weights and biases. But every course I get doesn't tell how the computed errors influence the weights. And the other issue is - if I perform back propagation for EACH training example ... again, how each neuron error influence the whole weights for my net (which are general for all training examples).

I will be grateful for your help

Prze Gee
  • 49
  • 1
  • 5

1 Answers1

0

While I am trying to get the deep learning flow, I can not find out one detail -> when I reach an error on every neuron (in backpropagations flow), what is next I should do with that all errors.

This isn't exactly correct, you have a loss function for your output, which is usually compared to some label. An example of this would be the classic MSE.

You have your error function and you're basically adjusting the weights of the different neurons according to the output error. If you take the derivative/gradient of error function and break down into the chain rule, you can see how to adjust each neuron. Look at this link for a detailed explanation - https://brilliant.org/wiki/backpropagation/.

And the other issue is - if I perform back propagation for EACH training >example ... again, how each neuron error influence the whole weights for my net >(which are general for all training examples).

Usually back propagation is performed per batch of examples. I reccommend reading up on stoachistic gradient descent to understand why the adjusting the neuron weights for batches is still valid for the whole training set.

ec2604
  • 501
  • 3
  • 11