Enhancement of Deep Learning Architectures
Dulani Meedeniya · 2023
This chapter discusses possible model enhancement techniques. Regularization is a technique that makes slight modifications to the deep learning model, to improve generalizability. Methods such as early stopping and dropout are discussed that enable the performance of the model. This chapter then explains data augmentation techniques, which refer to approaches for increasing the quantity of data by adding slightly changed copies of current data or creating new synthetic data from existing data. Next, the chapter discusses normalization, which is a data preprocessing technique for converting numerical data to a common scale without changing the form of the data. Recalling the hyperparameters discussed in Chapter 2 , this chapter discusses the tuning process of the learning rate, batch size, momentum, and weight decay. Moreover, neural architecture search is an approach to automatically find and design architectures that will yield models with optimal results on a specific task. The goal is to design the architecture using limited resources and with minimal human intervention. Finally, this chapter describes adversarial learning, which addresses the attacks on learning algorithms and the possible solutions.