Ahmed Fawzy Gad - Practical Computer Vision Applications Using Deep Learning with CNNs: With Detailed Examples in Python Using TensorFlow and Kivy

Any source code or other supplementary material referenced by the author in this book is available to readers on GitHub via the books product page, located at www.apress.com/9781484241660 . For more detailed information, please visit http://www.apress.com/source-code .

Artificial intelligence (AI for short) is the field of embedding human thinking into computers In other words, creating an artificial brain that mimics the functions of the biological brain. Whatever the human can do intelligently is now required to be moved into machines. First-generation AI focuses on problems that can be formally described by humans. Using AI, steps for doing something intelligent are described in a form of instructions that machines follow. Machines follow the human without changes. These features are characteristic of the first era of AI.

Humans can fully describe only simple problems such as Tic-Tac-Toe or even chess and fail to describe the more complicated problems. In chess, the problem can be simply explained by representing the board as a matrix of size 88, describing each piece and how it moves, and describing the goals. Machines will be restricted to those tasks formally described by humans. By programming such instructions, machines can play chess intelligently. Machine intelligence is now artificial. The machine itself is not intelligent, but humans have transferred their intelligence to the machine in the form of several static lines of code. By static, it is meant that the behavior is the same in all cases.

The machine, in this case, is tied to the human and cant work on its own. This is like a master-slave relationship. The human is the master and the machine is the slave, which just follows the humans orders and no more.

Embedding intelligent behavior inside chunks of code cant handle all intelligent behaviors of humans. Some simple tasks, such as sorting numbers or playing some games, can be described by humans and then handled by the machine with 100% of human intelligence. However, some complex tasks, such as speech-to-text, image recognition, sentiment analysis, and others, cant be solved by just code. Such problems could not be described by the human as done with chess. It is impossible to write code to recognize image objects such as cats. Such intelligent behavior of recognizing objects simply cant be solved using a static code because there is no single rule for classifying objects. There is no rule to recognize cats, for instance. Even if a rule is successfully created to recognize cats in one environment, it will definitely fail when applied in another. So how can we make machines intelligent in such tasks? This is machine learning (ML), in which rules are learned by machines.

To make the machine able to recognize objects, we can give it previous knowledge from experts in a way the machine can understand. Such knowledge-based systems form the second era of AI. One of the challenges in such systems is how to handle uncertainty and unknowns. Humans can recognize objects even in different and complex environments and are able to handle uncertainty and unknowns intelligently, but machines cant.

In ML, the human is responsible to do the complex task of investigating the data to find what types of features are able to categorize objects accurately. Unfortunately, it is a challenging task to find the best types of features to use. This is the question that researchers are trying to answer for different applications. For example, to diagnose a disease, the expert human starts by collecting data for both affected and nonaffected persons, labels such data well, and finds some types of features that are robust in discriminating between people with the disease and those without it. Such features may be age, gender, blood sugar, and blood pressure. This is a very challenging task because the larger the dataset, the more complex for humans to find features working across all samples.

These days, however, ML models can be trained to identify how to discriminate between the different classes. The ML algorithm is what finds the suitable mathematical function that creates the most robust relationship between the inputs and their outputs.

ML algorithms are not doing everything; the key intelligence is still found in the human expert, not in the machine. The human collects and labels the data, extracts the most suitable features, and selects the best ML algorithm. After that, the ML algorithm just learns what the human has told it. Still, the machine plays an important role in finding the rule by which the inputs are mapped to the outputs.

Usually, ML algorithms trained with data from a certain environment(s) cant work with other environments. This is a key limitation. There are huge amounts of data existing all over the world. Day after day, the data increases and traditional ML techniques are not suitable for its manipulation. For instance, images are complex to describe using a set of engineered features due to the variations even within the same environment. The work (i.e., feature engineering) should be repeated to make the ML algorithm suited to work with other environments.

If the human ability to find good discriminating features decreases as the number of classes increases, we can avoid depending on humans and leave that task for the machine. The machine itself will try to explore the data and find suitable features to discriminate the classes. Just give the machine the data and it will find what features to use in order to make a classifier. This is deep learning (DL). The convolutional neural network (CNN) DL model is the trend for working with large amounts of images.