How does a Deep Neural Network work?

Hi guys. I have a question, and i’ve been confused about this. How does Deep Neural Network works? Starting with training set and test set with the forward prop and backward prop. Cheers!

I think this thread could be interesting for you:

In case you have additional open points, please feel free to ask more specifically regarding the unclear steps.

Best regards

1 Like

Thanks for the post! So after i’ve read it, there’s some point that i still don’t get it. Here are the points:

  • So the network learns from the pattern it learns in training set?
  • If so, it predict using only Forward propagation and compute loss right?
  • But does it require a loop?

And this is my understanding about it, please do fix me if i’m wrong.
For me, basically it learns from training set, with forward propagation, compute loss, backward propagation, update parameters, and loop for num_iterations, then it test on test set with only forward propagation.

Thanks in advance!

The purpose of the iterations is to train and update the weights of your network (iteratively).

  • So you can incorporate more information (e.g. data w/ new labels) into the training process through iterations, so that you eventually can find a minimum of your cost function.
  • In fact you want to reach your global optimum. Through the iterations you would navigate from your net w/ initialised weights (usually high loss, not even close to an optimum) to a much better parametrization (= weights) which is much more close to the optimum, where your costs are minimised.

Oh thanks for your reply! Now, sorry if this would annoy you, but one more which is for you to correct me if i’m wrong based on my understanding. So here is my point based on what i undestand, please do tell me if something is wrong.

So, first You trained the dataset with forward propagation, compute the cost, backward propagation, and update the parameters. Then after trained the network for num_iteration times, you have the trained network. So now, it want to predict using the test sets by following the pattern it learns from the train sets. Then it predict with forward propagation and it doesn’t repeat by num_iteration, since the purpose of num_iterations is for train set. Is that correct?


Exactly! The steps are correct :+1:

But you train the model only with the training data.

Of course you can evaluate the performance in every iteration for both train and test set which give you the characteristic loss plots, see point 4)

1 Like

Oh, thanks for your respond!

This thread might be interesting for you as well:

Best regards