Appendix B. Minimally Qualified Readers and Deep Learning Basics

This appendix orients readers on who the book is for and the baseline skills needed. It targets practitioners and data scientists with intermediate Python proficiency—familiarity with data types, control flow, functions, classes, and installing/importing packages. A basic understanding of machine learning and deep learning concepts (such as losses, activation functions, and optimizers) is expected, but no prior PyTorch or generative AI experience is required. Early material introduces PyTorch through fundamentals and a hands-on end-to-end project, preparing readers to build generative models that create images, text, numbers, shapes, and audio.

The appendix then summarizes core deep learning ideas. Neural networks consist of input, hidden, and output layers; those with many hidden layers are considered deep. Common layer types include dense (fully connected) layers, convolutional layers that efficiently detect local spatial patterns with shared weights and translation invariance, and recurrent networks designed for sequences that carry information across time steps. Activation functions provide the nonlinearity that makes these models powerful: ReLU passes positive inputs, sigmoid maps values to 0–1 for binary decisions, softmax converts vectors into class probability distributions, and tanh outputs values between −1 and 1.

Finally, it outlines the training process: split data into train, validation, and test sets; iteratively update parameters to minimize a loss on the train set; monitor validation performance each epoch and stop when improvements stall to limit overfitting; and evaluate final generalization on the test set. It reviews key losses—MSE/L2 and MAE/L1 for regression, binary cross-entropy for binary classification, and categorical cross-entropy for multi-class—and explains how the learning rate governs convergence speed and stability. The discussion of optimizers spans mini-batch stochastic gradient descent and Adam, which adapts per-parameter learning rates from gradient moments and is effective for large models and datasets.

Figure B.1 The structure of a neural network. A neural network is composed of an input layer, zero, one, or more hidden layers, and an output layer. Each layer contains one or more neurons. Neurons in each layer are connected to those in the preceding and subsequent layers, with the strength of these connections represented by weights. In this figure, the neural network features an input layer with three neurons, two hidden layers with six and four neurons respectively, and an output layer with two neurons.

Figure B.2 How to train a neural network. The training dataset is divided into three subsets: the train set, the validation set, and the test set. The process for training a neural network involves the following steps. In the training phase, the train set is used to train the neural network and adjust its parameters to minimize the loss function. During each iteration (epoch) of training, the model updates its parameters based on data in the train set. In the validation phase of each iteration, the model is evaluated using the validation set. The performance on the validation set helps determine if the model is still improving. If the model's performance on the validation set continues to improve, the next iteration of training proceeds using the train set. If the model's performance on the validation set stops improving, the training process is stopped to prevent overfitting. Once training is complete, the trained model is evaluated on the test set. This evaluation provides the final testing results, giving an estimate of the model's performance on unseen data.

FAQ