AI Super-Simplified Part 1: What is Machine Learning?
Disclaimer: I am not an expert in this topic. This article is part of my commitment to the WPP AI Academy, which has requested students share what we learn. This article includes what I learned — in my own words — after taking the excellent course AI for Everyone by Andrew Ng.
WHY ALL THE HYPE?
A lot of what is called Artificial Intelligence (AI) falls under a giant set of technologies: automation, deep learning, neural networks, supervised learning, generative AI, NLP (Natural Language Processing) etc.
Of course, people hype everything — so anything that is close to those techniques will get called Artificial Intelligence. Let’s cut through the hype and see why tech is interesting and useful.
TYPES OF ARTIFICIAL INTELLIGENCE
AI can be broken down into a few general categories.
- AUTOMATION: Automating tasks that are repetitive. Recording macros in Excel, etc. running pre-configured reports daily, chat bots, voice assistants (Amazon Alexa, Google Home, Apple Siri)
- GENERATIVE: Generative is getting super-hyped at this moment. This is where software develops its own content based on user input. These apps are currently blowing up your news feed: AI-generated art, AI-generated chats, and AI-generated code.
- LEARNING: Learning involves feeding a huge set of data to a computer and requesting it notice patterns, if any. Humans then do things based on that information. There are many overlapping subcategories here (neural networks, machine learning, deep learning). Don’t be turned off by the terms — they are often used interchangeably. They have differences but like most jargon, you have to be clear how they are being used.
Here, we’re going to cover the Learning piece of AI. I think that is what is exciting and has practical use in the workplace right now. Automation is easy for people to understand — a computer does repetitive work, and people are familiar with it. Generative is easy to understand but not entirely applicable across the workplace — yet. IMHO, The topics closest to the average worker right now are data analysts and machine learning engineers.
LET’S MAKE A NEURAL NETWORK
So how does a machine, a computer, “learn?” It uses a neural network. A what? Let me show you. We’ll make one right now.
If I gave you a sequence of ten numbers like this: 1 1 1 1 1 1 4 1 1 1 and asked you to guess what number came next, what would you guess? Well, you’d probably guess 1. Why? Because 9 out of the 10 times it’s 1. Any clever-ish predictor of the future concludes — based on the given data — that there is a ninety percent chance the next number is 1.
Congrats. You built a neural network. You took data, found a pattern, and made a prediction algorithm. Based on the available data, you predicted the next number had a 90% chance of being 1. That’s what machine learning is. That model would be called a “neural network” if it was a piece of software.
MAKING A NEURAL NETWORK USEFUL
Let’s review a more useful example. Say you work at the local big box in gadget sales. You scan the receipts of people buying televisions and notice this.
What that data shows is that for every ten people who buy a TV, eight of them also purchase extension cords. A smart salesperson like you concludes that when someone purchases a television, you should also ask if they need extension cords. It’s not a guarantee, but there is an 80% chance of an additional sale. Store designers could also use this data, placing extension cords near the TVs, maybe even adding a sign, “Don’t forget extension cords!” etc.
Congrats. You built a second neural network! Based on available data, you made a prediction weighted with probability. You’ve probably seen this in action on online stores thousands of times. If you buy item A, the site will often try to also sell item B based on its data and probability. “People who bought item A also bought item B.”
RECAP: A NEURAL NETWORK DOES MACHINE LEARNING
The examples above are ultra-simple but intended to demonstrate what a neural network — often also called a model or algorithm — does. Recap: It takes a huge amount of data, and finds patterns (also called “learning”) then it can highlight — or help a human notice — certain things may happen in the future.
Of course real-life data in any field is going to be far more complex than these controlled examples. Think of sports — no matter which you enjoy, every sport is overflowing with data. Each player has a height, age, weight, percentage of successful action, and so much more. Then collectively each team has its own stats. It’s an ocean of information. Finding patterns based on that much numerical data is difficult for us biology-based machines. But that’s where computers excel.
We bio-brains have the edge in natural processing, such as whether visual data is a lamp or a charging lion. Computers are catching up in image recognition, but for now, they have the edge on data and we biologicals have the edge on real-world imagery.
WHY NEURAL NETWORKS ARE EXCITING
Since the dawn of humanity, we’ve wanted to predict the future. Now we can! Actually, not entirely. We have computers help us make predictions based on the provided data, and that data is weighted with probability. So a computer model can look at the massive amount of data on two sports teams and make a probability on who is going to win a game. But that probability isn’t a certainty, and is also dependent on what data it is fed.
Just look at this sample from NFL.com for football teams. And this is just for passing information, there are tables like this for all kinds of traits on the teams’ performance. How would you know which data is most important?
That’s where all the exciting careers come in. The human data engineer does the work of data analysis, collection, loading into the system. The human machine learning engineer build the software that crunches the numbers at speeds and volumes no human can match. Humans are still key in the process because they feed in the data, tell the computer which data to find a pattern on, and then take action based on that probability.
Notice I keep using phrases like “based on the provided data” and “weighted with probability.” An AI can’t anticipate the future, or some new sudden factor that changes things (such as an injured player, a drastic rise in extension cord prices, etc.) That’s why it’s important to remember the info from the neural network (or model, or algorithm) is subject to what data it is fed, and requires human analysis. And once more, it delivers a probability, not a certainty.
Garbage-In Garbage-Out. Bias-In Bias-Out. If your data is incomplete or contains flaws, those issues will be reflected in the model’s recommendations.
Data is everywhere now, and so are software engines, codes, etc. that help you build models as pieces of software. It’s a gold rush of sorts. There are many exciting developments but also lots of scams, hype, and confusion.
What is exciting is that there is no telling what patterns are hiding in the near-infinity of numbers that can be applied to the world. Wisely-assembled data and a well-designed algorithm will unlock patterns we never knew. Data analysis and machine learning teams can then advise actions to help exploit that hidden knowledge.
There is definitely an art to it all, but hopefully this explains a bit of the science and what the hype is about.
So now, chances are you’re nervous that is going to “replace” you. Is it?
Spoiler: No. Maybe. Depends.
Actually, that’s what we’ll discuss next time.
#neuralnetwork #deeplearning #machinelearning #artificialintelligence
Originally published at https://www.linkedin.com.
