Diving through the layers of Artificial Intelligence

When reading about artificial intelligence we see terms like neural networks, deep learning, machine learning, reinforcement learning and more thrown about. Some of these terms are often interrelated and can be substituted to discuss the same subject matter, whereas others are mutually exclusive. In this post, I aim to provide clarity on the various subgroups of AI to help you differentiate between various technologies/concepts.

Artificial Intelligence.

AI simply refers to any computer program that is capable of completing tasks that usually required human intelligence. This includes things like speech recognition, visual perception, and decision-making. There are many different disciplines for performing these tasks. Think about deciding whether you should introduce exercise or a new diet. Each method is different in the way they are executed but attempts to achieve the same outcome.

 

Level 1: Sub-streams of AI.

Symbolic AI was the dominant form of AI when it gained traction in the 1950s. This describes methods that are considered ‘human readable’ representations of various problems. These often take the form ‘if-then’ statements. For example, if my input is ‘a’ I should execute ‘x’, else if it is ‘b’ I should execute ‘y’ and so on. Given the complexities of real-world problems, a programmer would have specified many contexts and the software would become limited to the developer’s ability to consider various scenarios.

Queue Sub-Symbolic systems of AI. This stream is comprised of methods that do not require specific representations of knowledge to execute ‘intelligent’ functions. Due to the restrictive nature of Symbolic systems, Sub-Symbolic systems took precedence in new research in the 1980s and now is the dominant type of AI used today.

Level 2: Sub-symbolic AI.

There are many ways of creating a Sub-Symbolic AI, each of which is suited to different contexts. I list a few below:

• Behavioral Based AI: Often used in robotics which enables machines to learn from user-defined ‘mistakes’.
• Soft Computing: Uses approximations in solutions to computing hard tasks. Solutions are unpredictable and are between the values of 0 and 1. This ideal soft computing model replicates the human mind. This is the most popular methodology in AI and can be applied to fields varying from biology to management sciences.

Level 3: Specific Methodologies of Soft Computing

We will look at the specific constituents of Soft Computing since these are most relevant to topics covered in class.

• Machine Learning: “Gives the computer to learn without being explicitly programmed” – Arthur Samuel. The term ‘learning’ in machine learning refers to the existence of an objective function the computer is trying to optimize. By changing various parameters the system ‘learns’ what are more optimal weightings to assign to various points in the algorithm. A specific Machine Learning technique is Neural Networks, which as discussed in last week’s blog post, is comprised of multiple nodes in a network that pass information between each other.
• Evolutionary Computation: Inspired by ‘natural selection’. Sets of solutions are created by trial and error. Changing random variable re-iterates new solutions with the less desirable solutions being removed. This is very popular in structuring data and evolutionary biology.