Neural networks

Neural networks

par Elyes Ben Chaabane,
Number of replies: 2

Hello,

I have a question concerning neural networks.
How can we know which neural net is best ?

In this course we learned how to create custom NN, CNN, DNN and Linear neural nets (MLP), but there are others such as Recurrent Neural networks or The Long Short Term Memory Network, etc... .

Suppose I need to classify a bunch of data, are types of neural networks used by convention in some task or do I have to run all networks and test them one by one with a validation/training set ?

Thank you in advance,
Elyes.

In reply to Elyes Ben Chaabane

Re: Neural networks

par Sena Kiciroglu,
Hi Elyes,

Good question! These different neural networks were usually designed to tackle specific problems. For example, you already saw in class that we use CNNs when we are working with images. For sequential data (such as the coordinates of an object we are tracking) and NLP tasks, it's common to use RNNs, LSTMs or recently transformers. To generate images that look like paintings, we can use GANs. When we are dealing with inputs that look like graphs, we can use GCNs.

But of course, we can use a validation set to decide between types of networks if we are not sure which one would work best. And we use the validation set to determine the exact architecture (how many layers, how many nodes in the layers, etc.)

Best,
Sena