Why Deep Learning?

In this section, we will cover the importance of deep learning and the reasons for its popularity.

Deep learning is a subset of machine learning that uses multi-layer neural networks (large neural networks), inspired by the biological structure of the human brain, where neurons in a layer receive some input data, process it, and send the output to the following layer. These neural networks can consist of thousands of interconnected nodes (neurons), mostly organized in different layers, where one node is connected to several nodes in the previous layer from where it receives its input data, as well as being connected to several nodes in the following layer, to which it sends the output data after it has been processed.

Deep learning's popularity is due to its accuracy. It has achieved higher accuracy levels than other algorithms have ever before for complex data problems such as natural language processing (NLP). Deep learning's ability to perform outstandingly well has reached levels where machines can outperform humans, such as in the case of fraud detection. Deep learning models can not only optimize processes but also improve their quality. This has meant advances in revolutionary fields where accuracy is vital for safety reasons, such as self-driven cars.

Even though neural networks were theorized decades ago, there are two main reasons why they have recently become popular:

• Neural networks require, and actually capitalize on, vast amounts of labeled data to achieve an optimal solution. This means that for the algorithm to create an outstanding model, it requires hundreds of thousands or even millions of entries, containing both the features and the target values. For instance, as regards the image recognition of cats, the more images you have, the more features the model is able to detect, which makes it better.

  Note

  Labeled data refers to data that contains a set of features (characteristics that describe an instance) and a target value (the value to be achieved); for example, a dataset containing demographical and financial information, with a target feature that determines the wage of a person.

  The following plot shows the performance of deep learning against other algorithms in terms of the quantity of data:

Figure 1.1: Performance of deep learning against other algorithms

This is possible nowadays thanks to advances in software and hardware that allow us to gather and process such granularity.

• Neural networks require considerable computing power to be able to process such large amounts of data without taking weeks (or even longer) to be trained. Because the process of achieving the best possible model is based on trial and error, it is necessary to be able to run the training process as efficiently as possible.

  This can be achieved today through the use of GPUs, which can cut down the training time of a neural network from weeks to hours.

  Note

  With the objective of accelerating deep learning in order to be able to make use of large amounts of training data and construct state-of-the-art models, field-programmable gate arrays (FPGAs) and tensor processing units (TPUs) are being developed by major cloud computing providers, such as AWS, Microsoft Azure, and Google.

Applications of Deep Learning

Deep learning is revolutionizing technology and is already impacting our lives. Deep learning can be applied to a wide variety of situations, ranging from medical and safety (such as fraud detection) purposes to more trivial tasks, such as colorizing black and white images or translating text in real time.

Some of the applications of deep learning that are either under development or in use today include the following:

• Self-driving vehicles: Several companies, such as Google, have been working on the development of partially or totally self-driving vehicles that learn to drive by using digital sensors to identify the objects around them.
• Medical diagnosis: Deep learning is impacting this industry by improving the diagnosis accuracy of terminal diseases such as brain and breast cancer. This is done by classifying X-rays (or any other diagnostic imagery mechanisms) of new patients, based on labeled X-rays from previous patients that did or did not have cancer.
• Voice assistants: This may be one of the most popular applications nowadays, due to the proliferation of different voice-activated intelligent assistants, such as Apple's Siri, Google Home, and Amazon's Alexa.
• Automatic text generation: This means generating new text based on an input sentence. This is popularly used in email writing, where the email provider suggests the next couple of words to the user, based on the text that's already been written.
• Advertising: In the commercial world, deep learning is helping to increase the return on investment of advertising campaigns by targeting the right audiences and by creating more effective ads. One example of this is the generation of content in order to produce up-to-date and informative blogs that help to engage current customers and attract new ones.
• Price forecasting: For beginners, this is a typical example of what can be achieved through the use of machine learning algorithms. Price forecasting consists of training a model based on real data. For instance, in the field of real estate, this would consist of feeding a model with property characteristics and their final price in order to be able to predict the prices of future entries based solely on property characteristics.