AI and ML Terms in Plain English

Along with all the news on on Artificial Intelligence (AI) and Machine Learning (ML) come a lot of terms that many people may not understand. In this article I try to explain some of the ones that I get asked about most frequently.

The most common of course is the difference between AI and ML. These days the terms are pretty much used interchangeably by the public. Occasionally you'll see a pedantic article pointing out that they're not the same and should not be used in that way. That's because AI can really be thought of as a field that contains ML. Whereas ML are algorithms that learn from data, AI includes ML along with other techniques and algorithms such as optimization, simulations, genetic algorithms, rule based systems, Bayesian techniques, etc.

Typically Machine Learning is broken up into Supervised Learning, Unsupervised Learning and Reinforcement Learning. In Supervised Learning you start with data that has both features and ground truth answers. For example if you want to predict house prices you start with known features such as size of house, number of bedrooms, etc. for specific houses and known prices (answers) for those houses. This data is then processed through an ML algorithm (the training phase) which creates a model. The model can then process new feature data to make price predictions on previously unseen houses. The model does not exactly create rules but you can think of it as learning a large number of simple rules that it can then use to make predictions.

Supervised learning is often broken up into Classification and Regression which really just means you are trying to predict a label or a number. For example if you are trying to classify documents you may be trying to tell whether they primarily fall into the categories international news, sports, health, etc. That is you are trying to assign a label to the document. Regression, in this context, means you are trying to predict a number or score such as the reading level of a document or the price of a house.

In Unsupervised Learning you have the features but you don't have the label. The unsupervised learning algorithm processes the data to find and learn some kind of inherent structure. This is usually done with a simple task that can be judged from the data itself. For example, for text we may take a sentence, blank out a word, and try to get the model to guess what the missing word is. Or for other kinds of data we may pass the data through a network meant to compress it into a smaller (lossy) representation and then expand it back to the original data. In this case we learn the best lossy compressed representation of the original data. This compressed representation can then be used in other downstream tasks like classification, regression, generation or similarity search.

With Reinforcement Learning we conceptualize an agent affecting an environment and learning which decisions lead to the best outcomes. Here we don't have the "right answers" but the environment itself provides (delayed) feedback. Imagine a system playing a game where we don't know the best move to make but as the game rolls out the agent learns whether the move it made was good or not.

Neural Networks (NN), also called Deep Learning (DL), are a type of ML where we create (conceptual) neurons that have some information (weights) and connect to other neurons. Neurons only communicate with the neurons they are connected to and the weights affect the strength of the connection. In the learning phase the weights are adjusted so that error in the final answer is minimized. NNs can be used for most tasks including classification and regression. There are many kinds of NNs where the neurons are arranged in a particular pattern. These include feed forward, convolutional, recurrent, and transformers. The transformer architecture is relatively recent and is the basis for many of the systems being worked on today such as Large Languge Models and other generative AI systems.

Large Language Models (LLM) use transformer NNs to create a super smart auto-complete that has analyzed a lot of text. The size of the LLM and vast quantity of text leads to an emergent behavior where the LLM can be used to answer question as well as process new text it is given. This ability lets it be used to create summaries, extract information, restructure information along with many other useful tasks.

I hope this was useful. If you'd like me to explain any other term get in touch

Want to get notified of new articles and insights?