For the purpose of this book, consider artificial intelligence (AI) as the field of computer science responsible for making agents (software/robots) that act to solve a specific problem. In this case, "intelligent" means that the agent is flexible and it perceives its environment through sensors and will take actions that maximize its chances to succeed at some particular goal.

We want an AI to maximize something that is named Expected Utility or the probability of getting some sort of satisfaction by doing an action. An easy to understand example of this is by going to school, you will maximize your expected utility of getting a job.

AI aspires to replace the error-prone human intelligence involved in completing tedious everyday tasks. Some central components of human intelligence that AI aims to mimic (and an intelligent agent should have) are:

• Natural Language...

Very vaguely inspired by the biological network of neurons residing in our brain, artificial neural networks (ANNs) are made up of a collection of units named artificial neurons that are organized into the following three types of layers:

• Input layer
• Hidden layer
• Output layer

The basic artificial neuron works (see the following image) by calculating a dot product between an input and its internal weights, and the results is then passed to a nonlinear activation function f (sigmoid, in this example). These artificial neurons are then connected together to form a network. During the training of this network, the aim is to find the proper set of weights that will help with whatever task we want our network to do:

Next, we have an example of a 2-layer feed forward artificial neural network. Imagine that the connections between neurons are the weights...

We will now look at another type of neural network that is especially designed to work with data that has some spatial properties, such as images. This type of neural network is called a Convolutional Neural Network (CNN).

A CNN is mainly composed of layers called convolution layers that filter their layer inputs to find useful features within those inputs. This filtering operation is called convolution, which gives rise to the name of this kind of neural network.

The following diagram shows the 2-D convolution operation on an image and its result. It is important to remember that the filter kernel has a depth that matches the depth of the input (3 in this case):

It is also important to be clear that an input to a convolution layer doesn't have to be a 1 or 3 channel image. Input tensors to a convolution layer can have any amount of channels...

Before we start, there is a bit of good news: using TensorFlow, you don't need to take care about writing backpropagation or gradient descent code and also all common types of layers are already implemented, so things should be easier.

In the TensorFlow example here, we will change things a bit from what you learned in Chapter 1, Setup and Introduction to TensorFlow, and use the tf.layers API to create whole layers of our network with ease:

import tensorflow as tf 
from tensorflow.examples.tutorials.mnist import input_data 
mnist = input_data.read_data_sets("MNIST_data/", one_hot=True) 
# MNIST data input (img shape: 28*28) 
num_input = 28*28*1 
# MNIST total classes (0-9 digits) 
num_classes = 10 
 
# Define model I/O (Placeholders are used to send/get information from graph) 
x_ = tf.placeholder("float", shape=[None...

In this chapter, we introduced you to ML and artificial intelligence. We looked at what artificial neural networks are and how to train them. After this, we looked at CNNs and their main building blocks. We explained how to use TensorFlow to train your own CNN for recognizing digits. Finally, we had an introduction to Tensorboard and saw how it can be used to help visualize important statistics while training models in TensorFlow.

In the next chapter, we are going to look more closely at the task of image classification and how we can use CNNs and TensorFlow to solve this task.