Overview

The current technology has the power to reshape society once more. The recent developments in Artificial Intelligence will directly impact our daily lives and several studies indicate AI will have a great impact on jobs. Like previous industrial revolutions job positions will cease to exist and others will be created. A report from the World Economic Forum says 1.4 million US jobs will be automated before 2026 and they believe the re-education is the solution to the impacted working force.

What it is?

Deep learning is the field that studies any machine learning algorithm that learns multiple levels of representation (abstraction). Although this definition is very generic most successful solutions are entirely based in neural networks. Truthfully almost everything in deep learning was built around neural networks and there is not much work with other machine learning algorithms. The deep learning definition is still very controversial and since it is still an area under development more time is required until a precise definition is globally accepted.

Deep learning or deep neural networks are machine learning algorithms that require a dataset. In turn, a dataset is a set of data examples and annotations (labels or ground truth) that must be carefully selected to represent a problem. These annotations contain information that the model must learn or, precisely, annotations are exact representations that the model should output for a given input. The annotation requirement place these models in the supervised learning subcategory from machine learning.

Although we can use the technology we can hardly explain how it works. The common approach right now is based on good practices followed by fine tuning (adjusting) the model. We have guides to teach what to do and network models exhaustively tested in several datasets. However, there is no mathematical description of deep neural networks internal computations. These computations are discovered during a training process that uses the dataset. With a better understanding we could be able to manually create networks for each problem selecting the number of layers and directly adjusting the weights without the need of training datasets.

What it can do?

From new products and services to new types of computer virus the list of possible AI applications is vast. All major tech companies are heavily investing on AI powered products. The most famous products are self-driving cars, search engines, recommendation systems, personal assistants (Alexa, Cortana, Google assistant, Siri) and domestic robots for household chores.

Unfortunately this technology is not only being used for good. Like the human intelligence, deep learning is dual-use and can be applied towards beneficial and harmful ends. Also the same tools used to create AI products can also be used to create malicious applications. The list of these malicious applications is also vast and include fake data generation (news, audio, image, videos), phishing attacks, autonomous weapon systems, software vulnerability discovery, hacking systems and surveillance platforms. For more information on how AI can be maliciously used see: malicious AI report.

Self-driving cars are one of the main examples of AI powered products. All major car manufactures are planning to release autonomous cars between 2019 and 2021. The first impact of this technology is the replacement of taxi drivers and truck drivers. Companies like Uber even have their own research team dedicated to self-driving cars.

Future perspectives

AI will challenge our perception of reality. Deep learning is able to generate realistic images and audio and they will likely revolutionize the audiovisual and gaming industry. But most important we will have this power at our hands and we will be able to modify our surroundings generating fake images and audio. Once this happens we will certainly face situations were it will not be possible to recognize what is real.