For the rest of this chapter, we will focus on using deep neural networks to make a prediction. When compared to using traditional multiple linear regression, you'll be pleased to find that our neural network has the following advantages:

• We don't need to select or screen features. Neural networks are amazing feature engineering machines and can learn what features are relevant and ignore the irrelevant ones.
• Given an adequately complex network, feature interactions are also learned (for example, the effect of x₁ * x₂ in addition to the independent effects of x₁ and x₂)
• As you have maybe guessed by now, we can also learn higher order polynomial relationships (for example, x₂³ )
• And lastly, we don't have to constrain ourselves to only modelling normal distributions, or using different models for non-normal distributions, so long as we make sure our final activation can possibly model the distribution.