Deep Learning (DL) DL is a one kind of representation learning. For example, think about detecting objects from images. One node in a hidden layer (HL) may detect the rectangle pattern from the images and used that information for subsequence layers to detect a checkerboard inside the rectangle and later it uses this pattern to find something like a cat and so on. So, it goes like a representation learning.

Activation Function (AF) AF tells you how you want to adjust the output produced by each node in NN. For example, RELU that uses tanh (np.tanh()) function: always returns 0 if input is negative otherwise gives identical or same with the input of any positive numbers)

Loss Function (LF) LF computes the error at the final output nodes by comparing predicted final NN outputs with actual target outputs. For example, mean squared error (mse) computes a scalar error value (i.e., squared root of summing all errors at the output layer).

Building a mode for regression problems (e.g., housing price prediction and stock prediction) mse is a good choice as a loss function

Another loss function categorical_crossentropy is used for classification problem

Model Building Steps Specify architecture Compile Fit Predict