Building Intuition: What is AI/ML?

TL;DR: AI/ML = ways to find patterns in data

LLMs have once again bubbled artificial intelligence (AI) and machine learning (ML) to the forefront of technology. Machine learning is the subset of the broader category of AI. In terms of industry, the term is often used interchangeably with AI, and Big Data.

Why? Because at the end of the day, there are only three real concerns with machine learning as a part of a product. One is the algorithmic side (in neural nets, this is the architecture). Two is the data side; the quality, quantity, and relevance of the data all matter. Three is deploying, serving, and operations.

Industry focus has shifted from rule-based AI to data based AI — aka Machine Learning. The scope of ML encompasses all software programs that can learn from data. In the frame I laid out above, this combines the data and the algorithms. There are many formulaic machine learning algorithms such as K-Means, K Nearest Neighbors, or Linear Regression.

Then there are neural networks. Neural networks are the power house of machine learning. They “algorithmic” side of machine learning nowadays is mostly around the structure of the neural network. At the time of writing, transformer models dominate both the language and vision sides of machine learning.

How Do Machines Learn from Data?

There are more or less two types of machine learning: supervised, and unsupervised. Technically, there are also many other types including transfer learning, reinforcement learning, semi-supervised learning, and more. Really, these methods are all just different ways of combining supervised and unsupervised learning.

Supervised learning describes giving an input-output combination for the machine learning model to work on. Unsupervised learning describes giving a model a set of data and letting it work out as it’s problem formulation and architecture dictate.

At the beginning, models are all initialized with random weights; you can think of this as a random algorithm or transformation. As you train the model, it updates these weights. Check out how to build a neural network from scratch for a mathematical deep dive.

Since a model updates the weights as it processes the data, eventually the model “learns” the data. For supervised learning, the model learns how to map the inputs to the outputs. For unsupervised learning, it finds the patterns in the data that are relevant to the problem at hand.

What does it all mean?

For someone who isn’t working directly on AI/ML, you can think of AI/ML as any or all of the following depending on your flavor:

• We don’t know how the data maps to the output, but we used a machine to find existing patterns
• We want to make something that can make some decisions but don’t know the rules for making those decisions, so we asked a machine to find some rules to follow. Unfortunately, the machine can’t explain those rules to us
• We need to raise money for our startup and want to ride the wave (at least from 2018–2023 this holds true, we will see what happens going down the line)
• We know there are patterns in the data, but we don’t know what they are
• We think the data should map to an output that we want, but we don’t know if it does or how to do that, so maybe a software program can find that for us