Visual Introduction to Deep Learning

by Meor Amer

From the introduction

By the end of this book, you will be able to build a visual intuition about deep learning and neural networks.

This book doesn’t cover mathematical proofs and code examples. As you advance
your learning further, these are the domains you should progress into. They will
provide you with the depth you need to be successful in this field.

Table of contents

- Introduction
- Foundations

1. Linear regression
2. Non-linear regression
3. Binary classification
4. Multi-class classification

- The Bigger Picture

236 pages, ebook

Published January 1, 2022

About the author

Meor Amer's mission is to help create data-driven professionals and youths via an enjoyable learning experience. He has previously worked with clients in over fifteen countries for deploying telecommunications data analytics solutions and running training & enablement programs.

Reviews

[Alejandro Teruel · Rating 4 out of 5]

An intriguing effort to explain, visually, how neural networks work. He starts by showing how a single neuron can be used to propose a linear regression for a set of data points. The data points are divided into training points and points that are reserved to validate the proposed linear regression. Each data point consists of the value of a feature and the correct result. In the following chapters, he shows how a simple neural network can be used to propose non-linear regressions, and carry out binary classifications, and multi-class classifications. This progression allows him to gently introduce the neural network's Predict-Measure-Feedback-Adjust cycle, the notion, and use of a neuron's inputs, outputs, and parameters (weights and biases), the relationship between the weighted sum of inputs into each neuron and its activation function -showing the effect of using the Rectified Linear Unit, the sigmoid, and the softmax functions as activation functions-, the loss function -showing how the mean squared error, the binary cross, and the categorical cross-entropy functions can be used-, and how the loss function derivative is used to adjust each neuron's parameters. At the end of the chapter on multi-class classification, Amer skims over hyper-parameters such as the size of the network, the proper choice of activation and loss functions, the learning rate (alpha), the number of epochs (cycles), and input data batch sizes, without explaining how to choose or adjust them.

This all sounds daunting, but Amer's picture-based approach nicely manages to develop the reader's intuition without delving into the more detailed mathematical details.

In the last chapter, The bigger picture, the author then shows how some basic changes to neural network architectures such as convoluted neural, generative adversarial, and recurrent networks makes them able to recognize images or objects or carry out image segmentation, identify sentiments in text, translate texts and generate text or images. I felt this chapter was all too brief and would have liked Meor Amer to go into more detailed explanations, particularly with recurrent networks.

The book only requires some elementary mathematical knowledge about functions and the notion of derivatives. By the end of the book, the reader will have a general idea about the elements that make up neural networks, and how such networks work. The book is an excellent introduction to such AI models and will leave the book with more pointed questions about how and why such models work and be motivated and better prepared to move on to more technical books on the subject. Amer ends by providing a short list of recommended further readings.

As someone interested in understanding how neural networks and deep learning work, but who has not worked on linear algebra and basic calculus for many years, this proved a tantalizing glimpse of the area and I look forward to exploring the subject further.