With this high-level guide, I want to clarify ideas about terms that involve terms such as Artificial Intelligence and Machine Learning and what their applications and potential are. You will also be introduced to the Jargon of these incredible themes in continuous evolution. You will understand the hierarchy and contexts in which to read any news or fact that mentions AI or Machine Learning.
The term Machine Learning is often confused with the term Artificial Intelligence (AI), which however refers to a set of techniques, methodologies, and algorithms that includes those of Machine Learning (abbreviated to ML). The objective of the AI field is to create "intelligent" agents, which can have a perception of the real world through sensors, make decisions thanks to their "intelligence", and finally implement them through actuators. If you want to make an analogy with a human being, you can think of sensors as the sense organs (eyes, ears...), the intelligence is that of our brain and the actuators can be the hands, feet or more generally the muscles. However, the biological analogy is limited, since it is not necessary for the sensors of an agent to detect physical phenomena, such as temperature or time, but it could detect data, relations, or inputs of any other kind. Machine learning is generally considered as a subset of the AI field. Specifically, the term Machine Learning means the science of programming a machine so that it can learn to perform a task from the data, rather than through programming. Deep Learning is a subset of the Machine Learning field a further ensemble of techniques and algorithms that exploit neural networks for the solution of problems such as Computer Vision or Natural Language Understanding.
A more engineering definition of Machine Learning is that given by Tom Mitchell: "A machine is learning from experience E with respect to task T and a metric of P, if its performance on T, measured with P, improves with experience E."
This theoretical classification may be difficult to distinguish in practice, but it becomes clearer with an example, such as the autonomous driving system of a new generation car. This can, in fact, be considered as a whole an AI system: it collects data from the real world (road, cars, road signs, speed) through sensors, possesses a set of "rules" of behavior that constitutes its intelligence ("if you are about to collide with an object, brake preventively") and possesses a set of actuators through which it implements decisions (brakes, steering, ABS). In this example, the AI system of the car has a sub-module of Machine Learning that allows it to improve over time, as the car "sees" different roads and situations, and increases its performance compared to some metric (safety, reaction time to braking, consumption). This Machine Learning module is then made up of further sub-modules, among which there can be a Deep Learning module that specifically deals with the vision that the car has of the real world, through a neural network trained to recognize the route of the road, the shapes of cars, distinguish them from those of trees and pedestrians, and so on.
In computer science, the classic approach to the problem consists in designing a process (algorithm) that is then realized in a programming language (implementation), thus solving the problem in a deterministic way. This determinism can be identified with the ability to consciously interpret each step of the algorithm and understand what action is taken at each step and why. It is important to note that I do not need examples of input data to produce the output considered correct.
The approach of Machine Learning, at an intuitive level, consists instead in "seeing" a lot of input examples, each matched with the output considered correct, and programming the machine to detect the recurring patterns that occur in the data. These patterns will then be used by the machine as a "knowledge base", to try to return the correct output, if a new input is provided (never seen before). The process of "training" the machine is called training, and its result is a model (a set of parameters properly "adjusted" so that they recognize the patterns that interest us for the given problem). Once a model is obtained, it can be used to assign the correct output to inputs that it has never "seen". As it arises spontaneously to think, for a model it is determinant the ability to generalize starting from the examples on which it has been trained, otherwise, it would need to see every possible input that could be presented to it in input!
Let's see some practical examples of how Machine Learning is applied in the real world.
It is an application of Machine Learning:
-
An "intelligent" spam email filter, which the more spam emails it processes, the more effective it becomes in recognizing and distinguishing them from useful emails.
-
A recognizer of hand-drawn figures or symbols, such as numbers, letters, or figures.
-
A system of customization of content suggested based on what you've already seen (Netflix that recommends the next movie to watch, Facebook that suggests groups that might interest you, Amazon that shows products often purchased together).
-
A sentiment analysis system capable of extracting important information from tweets or other data generated by humans about a particular event or character. For example, you can do a sentimental analysis of the "Donald Trump" concept during elections, and understand how it is perceived by people who post tweets about it (negative, positive, neutral...).
-
A system of writing assistance that, as the text is composed, foresees the next words and suggests them. Examples of this system are extremely popular, such as the automatic completion of smartphones or the Facebook search bar.
-
A Computer Vision system that recognizes the model and brand of a car by "seeing" it with a sensor (typically a video camera).
We also see some counter-examples.
It is not a Machine Learning application:
-
A search engine based on keywords: it does not improve with experience, and to make it more "intelligent" you have to program it explicitly.
-
A customer service interface that has pre-programmed responses to specific needs. For example, if I detect the word "ORDER" and an alphanumeric code in the customer's request message, then I show the status of the shipment identified by that code.
-
A sensor programmed to detect car number plates (e.g. the Safety Tutor system). Although it "sees" license plates and can, therefore, be confused with the concept of Computer Vision, it does not improve with experience, and its performance is determined solely by its programming at the design stage, where it is explicitly calibrated to detect a standard object (the license plate).
-
The famous mobile application Akinator, which tries to guess which historical or cinematic character the user is thinking about. Through a series of targeted questions (does he have hair? is he an alien? appears in Star Wars?) he explores a database of characters and by asking questions, by exclusion, he arrives at the answer. This, although it may seem like an "intelligent" program, is not really: it doesn't learn by improving with experience without being programmed to do so and follows a deterministic process easily explained by a human being.
Imagine having to create a spam e-mail filter with traditional programming techniques:
-
First, we would try to understand what the typical spam e-mail looks like. We might notice that some words or phrases (for example "For Yourself", "Occasion", "Free only for today!") tend to appear often in annoying spam emails. Maybe we could notice further patterns in other parts of the mail, for example in the string representing the sender's mailbox, or that several emails are received from the same mailboxes in a few days.
-
We would write an algorithm to detect these patterns that we have noticed, for example using regular expressions to detect sequences of "suspicious" words and the algorithm would report as spam an email that contains a number of these patterns.
-
We would test the program and repeat steps 1 and 2 until we get a satisfactory version that signals malicious e-mails fairly accurately.
From the moment the task is not trivial, and those who create spam emails reinvent themselves every day to escape the control of the filters (fooling, for example, the ML anti-spam model of Gmail), the program would very soon become a huge set of complex "hard-coded" rules, chaotic and very difficult to maintain. Its main limit, however, is that to remain effective day after day, it must be updated daily and rules on rules must be added by hand!
Instead, a spam filter consisting of a Machine Learning model would automatically learn which sentences, words, and patterns are valid indicators of spam danger, updating itself every time a new spam mail is recognized. The program is much shorter, easy to maintain and accurate in recognition.
In summary, Machine Learning techniques are appropriate for:
-
Fluid environments, where the problem changes constantly and in an unpredictable way (a mountain road never "seen" before, an email containing words never encountered, an unknown pattern in the peaks of users in a website).
-
Complex problems where the traditional approach would be impracticable or would lead to "Frankenstein" of code with high maintenance cost. The Machine Learning approach can greatly simplify the code and achieve better performance.
-
Get information about complex problems and discover patterns through large amounts of data.
Now you should be more comfortable with high-level ideas of the concepts of AI and ML, and every time you see a term you'll know how to contextualize it. You will also see how the terms are crippled and misused, and maybe you could recommend this guide to the next person to clarify his ideas! If you like Virgil's content and share our vision of open education accessible to all, share and register in the Google Form, you'll be updated every time a new guide comes out!