Print bookPrint book

2.2 Types of AI

Site: Wolearn: Openness, Collaboration, Innovation
Course: AI Essential
Book: 2.2 Types of AI
Printed by: Guest user
Date: Wednesday, 8 December 2021, 4:58 AM

Table of contents

1. Types of AI

In this video, David discusses the different types of AI, and in particular how "Good Old Fashioned AI" is related to machine learning, and recent advances in deep learning.



In this course we are interested in Weak AI, sometimes called Narrow AI (as weak is somewhat of a misnomer given how powerful AI can be). Later we will see how organisations are using both GOFAI approaches, using rules and templates, as well as machine learning techniques, including deep learning networks.

The video has introduced some important terms that are worth remembering:

  • Weak vs Strong AI

Weak AI - AI focusing on solving a very specific problem in a narrow domain, such as facial or speach recognition, expert systems, or game-playing. Real and available today.
Strong AI - AI that aims to replicate the intelligence of a human-being, and their ability to solve general problems. Mostly for science-fiction writers and philosophers.  

  • Good Old Fashioned AI

GOFAI - A term used to describe symbolic approaches to AI, typically in the form of rules and logic. GOFAI typically encodes existing knowledge (for example, in an expert system) that can then be queried.

  • Supervised vs Unsupervised Learning 

Machine Learning - A term used to describe AI approaches that learn the rules themselves, by creating a statistical model based on data. Machine learning techniques can be:

Supervised - typically used to make predictions. In this case the model is trained using a data set with a known output variable. For example, we could look at website shopping activity and train a machine learner using the ultimate choice to purchase or not purchase as the known output variable. Once trained the algorithm can use the model to look at a new shopper it has not seen before, and predict whether they were going to make a purchase.

Unsupervised - typically used to better understand a dataset. In this case the model is trained using a dataset but with no known output variable. For example, we could analyse the characteristics of existing customers in order to classify them into different types.

  • Neural Networks and Deep Learning

Neural Networks - are a type of Machine Learning approach where inputs propagate through a network, altered by different weights between the nodes, and finally reappearing as outputs on the other side. There are different algorithms for training the network (setting all the weights) based on data.

Deep Learning - is an approach to Neural Networks where there are many hidden layers between the inputs and outputs. Deep Networks are difficult to train, and require significant computing power, but are capable of detecting much more complex patterns than shallow neural networks.