For more than a decade, hundreds of millions of people on planet Earth have, unbeknownst to them, interacted on a daily basis with the most advanced artificial intelligence (AI) ever created. The AI I am referring to is none other than the Google search engine. Surprised? Google’s online search engine actually meets the definition of an AI system perfectly; I’ll come back to the definition of AI later in this article.
Regardless of the fact that an AI has served humanity so well for so long, only this past year the term artificial intelligence has become mainstream and everyone, expert or not, has felt it necessary to share their opinions about it. Mainstream thinking on AI today is that robots, i.e., realizations of AI systems, are going to take away our jobs, act without morals and destroy humanity unless we figure out a way to control them. It is very likely that robots won’t do any of the latter but it is better for business to catastrophize than to be reasonable. Anyway, this article is not about futurologists making, yet again, inaccurate predictions about another major threat to humanity and its fast approaching doomsday.
This article is an attempt at clarifying the difference between the terms artificial intelligence (AI), machine learning (ML), and deep learning (DL). These three terms have been abused greatly in mass media, online and offline, by being used as synonyms. They are not, so let me explain.
The actual relationship among these three terms is that deep learning is a type of machine learning and machine learning is one component of an artificial intelligence. That is, AI is the more general term and DL the more specific term.
Now, one often hears that there is no agreed upon definition for AI and, even though this might be true, it is well understood within the academic community that AI is a broader concept than ML and, of course, DL. So what is AI? The definition I prefer is the one given in the most excellent textbook, Artificial Intelligence: Foundations of Computational Agents, and it is the following,
Artificial intelligence, or AI, is the field that studies the synthesis and analysis of computational agents that act intelligently.
Obviously this definition is rather general and many terms must be explained further. The book goes on to discuss in detail the concept of agency, what makes a computational agent and what it means to act intelligently. In addition, the scientific and engineering goals of AI are laid out in clear language. If you want to know the details, consider reading the related book chapter from the book’s official online copy here. For now, I want to focus our attention on the “acts intelligently” attribute of AI agents.
To act intelligently, we are told, an agent (software, e.g., a web crawler, or hardware, e.g., a robot) must take action appropriate to its circumstances, goals, and perceptual as well as computational limitations. Furthermore, this agent should learn from experience. This is where machine learning comes into artificial intelligence. Machine Learning is the discipline focused on extracting knowledge from data and using it to improve agent performance on some task. There are many different types of machine learning such as supervised and unsupervised, online and offline, logic-based and probabilistic (or combinations of the two), as well as reinforcement learning. In a future article I will provide an overview of machine learning. For the time being, I just want to emphasize that machine learning is only one part of what makes an AI.
Now that we understand the relationship between machine learning and artificial intelligence, what about deep learning?
As the name suggests, deep learning is a type of machine learning. It is difficult to fully explain deep learning in a few sentences but I can briefly say that it is a method of function approximation via the composition of a large number of linear and non-linear functions organized hierarchically. This hierarchy is composed of layers of sets of these linear functions. Deep learning architectures are the well known artificial neural networks (ANN) from decades ago with the main difference being that today the number of layers in the hierarchy is much larger, i.e., deeper, than in past ANNs. Access to vast amounts of data and much faster computers, especially specialized hardware such as GPUs, have allowed ANNs to make a comeback since falling out of favor in the late 80s and 90s. Deep learning systems have outperformed other types of machine learning systems on a number of difficult problems such as speech and image recognition. This explains their popularity skyrocketing in the last few years. However, deep learning has not totally eclipsed the need for all other machine learning algorithms as the media would have you believe. One should always remember the “no free lunch” theorem(s) that basically states that no one algorithm is best suited for solving all learning problems. Machine learning still requires that an expert spends a considerable amount of time and effort to select the best performing algorithm to solve a given problem.
My goal with this article was to clarify the difference between artificial intelligence, machine learning, and deep learning in order to rescue you from the incredible amount of misinformation available online today. I hope I have succeeded!