Convolutional network related

Hi @Manul_Das ,

The decision of which features are represented by each layer are decided automatically by the filters applied during each convolution.

So you may ask next: Who/how are the filters defined? (at least I asked that myself). The answer is: The filters are learned by the model during the training phase. We only have to define the structure of the filters, for example we define if a filter is a 3x3 or a 5x5 filter. Once we define this and the training starts, each filter will learn what patterns to represent.

Finally, when we design the CNN, we define multiple filters along the model. These filters will learn which features to represent, and it is understood that the initial filters are more abstract or more oriented to detect borders, edges, corners, while the filters towards the end are usually more oriented to represent more concrete details of the data.

I hope this sheds some light to your question. Please share any additional question.

Juan