Model Regression to Neural Networks

I may have missed something. In the previous course, we had specific training data and calculated the cost functions to generate our models. Then in this (2nd) course, we are programming neural networks. It seems like we are simply feeding numbers into the neural network. But how does the neural network get trained, and know how to split the work up between nodes?

It might help if you can clarify a bit what you mean by “the previous course”. Here in DLS Course 1, we first learn about Logistic Regression and then “feed forward fully connected” Neural Networks. In both cases, we are doing what is called “supervised learning”: that means we have labelled training data. For each input sample of whatever the data is (e.g. an image), we also have an associated “label” that tells us the correct answer for that data sample. The problem that we use as the example here in DLS C1 is taking 64 x 64 RGB images and asking the algorithm whether each picture contains a cat or not. So the labels are “true” (1) for images that contain cats and “false” (0) otherwise.

Then Prof Ng tells us how training is accomplished: we have a “loss function” which gives us a concrete metric for whether the answer (prediction) produced by our network is correct or not. The “magic” is that it turns out we can use the derivatives (gradients) of the loss function to push the parameters of the network (the weight and bias values in all the layers) towards better and better answers. That process is called “back propagation”. We simply have to choose the size and shape of the network and then run the training and find out what happens. We don’t need to tell the training how to split up the work between nodes: it figures that out automatically through back propagation. Of course it doesn’t always work if you don’t choose an appropriate architecture given the complexity of the problem you need to solve. Prof Ng will spend quite a bit of time in Course 2 of DLS helping us understand the process for choosing network architectures and then for deciding how they need to be changed in cases that we don’t get good enough results.

So we are just at the beginning of the learning process here. Please stay tuned for all that we will learn here in C1 and in the later courses in DLS as well.

Oh, sorry. I’m in the “Advanced Algorithms” course now, end of week 1.

Which specialization is that a part of? Note that you posted this under DLS Course 1, so I assumed that was what you were asking about.

Machine Learning. Courses:
[Supervised Machine Learning: Regression and Classification]
Coursera | Online Courses & Credentials From Top Educators. Join for Free | Coursera)
[Advanced Learning Algorithms]
(Coursera | Online Courses & Credentials From Top Educators. Join for Free | Coursera)
[Unsupervised Learning, Recommenders, Reinforcement Learning]
Coursera | Online Courses & Credentials From Top Educators. Join for Free | Coursera)

I just took a quick look at the syllabus and I think they are just talking about Neural Networks. So I think what I said in my earlier reply does apply.

OK, thanks.

I believe this thread is a duplicate of another post.

Closing it.