What is AI, Really?
Program, math, artificial brain? No, it's a …
The other day I was chatting with a CIO of a multi-national corporation; a friend of mine, and he asked me to explain to him how AI works? I went on to do a Google search and to my surprise, there isn’t anything relevant. You can find everything that talks about learning algorithms, applications of AI, python and even the Gartner Magic Quadrant report. There is nothing that explains AI in simple terms; something that a CIO of a multi-national corporation could quickly understand.
That leads me to think. The term AI, a short-form for Artificial Intelligence is a buzzword of the century. We have all heard of AI, but do we really know what’s under the hood? AI is complex. AI is a geeky thing. AI is something a software engineer creates. I am not a software engineer, so I don’t understand AI; is something you might say.
This is not true. At its core principles, AI is really easy to understand. In the next few minutes, I am going to simplify AI for you. We are going to make it so simple, that at the end of it, you will know AI as much as a software engineer. Maybe, understand it even better than some software engineers do.
I have been doing AI since 2004. Most of you have probably heard of AI only within the last 5 years. But AI has existed since World World II. Alan Turing made the worlds first AI machine to crack the Enigma code. If you are not aware of this, then I would recommend watching “The Imitation Game” movie. It is a brilliant portrayal of what AI looked like in the early days. This was even before computers were born. So if you believe AI is a computer program, you are only partially right. The world used AI systems even before computers or software was invented.
I am going to try and simplify AI for you. Together we will break it down to its core principles. Simply follow along, and hopefully, I will be able to offer you a very deep understanding of AI. Something that you would have never known off before.
Alan Turing’s machine proves that we can have AI without modern-day computers. We use computers today for AI, as it is easy to do so. Underneath, AI is just a simple logical system that mimics human intelligence. Yes, that is correct. It mimics. When any system exhibits a behaviour that is supposedly a very specific trait of intelligent living beings, we call such a system to be artificially intelligent.
You see, humans are truly intelligent. An AI system although mimics intelligence is not really intelligent in any form. Modern AI applications are nothing but simple math and logic conditions. When we trigger this math and logic, it produces an output that appears to be intelligent.
Enough of the jargon now. Let us look at what these math + logic conditions actually do. I will start with taking an example of an elastic hairband. We all know hairbands become loose after repeated use. In the world of Machine Learning, we would say that the hairbands are learning what you want them to be, than what they currently are. Where in, when you are stretching them you want them to be a little stretched, instead of what they are in their resting place. When you keep stretching over and over again, we can say that the hairband learns what you are doing, and eventually learns to remain stretched.
All AI algorithms work in the same way. They have 2 phases. The first phase is the learning phase, and the second is the predict phase. All AI models start with a default state, just like the hairband. A Data Scientist then passes data to the model to train the model. This action is the same as stretching the hairband. The model does not learn at first, it shows no impact. So what does the Data Scientist keep doing? Passes more data. Keeps passing data over and over again. Same as keeping on stretching the band over and over again. When you have passed enough data, the AI system has moulded itself enough to become similar to what the data intended it to be. This is the same as the hairband becoming loose; because when we were stretching it, we were trying to make it loose. Do it over and over again, the hairband finally gives way and becomes loose.
The first phase of training an AI algorithm on data is also commonly known as Machine Learning (ML). Keep in mind that machine learning and AI are not completely different things. ML is a 100% subset of AI. Everything we do in ML is a part of AI. But everything in AI is not ML. Such as using a trained machine learning model is not ML. The process of training the model is ML.
Now there are several different types of AI models. It can be Decision Trees, Support Vector Machines, Gradient Boosting Machines, Neural Networks, or even Deep Learning Neural Networks. All of these models are statistical systems. They have different mathematical equations, and these equations can learn different things faster than others.
Yes, I understand that the space of AI is more complex than a simple hairband example. You might even raise a question if AI is that simple? The answer is YES, it is that simple. At a basic level, all AI systems have a mathematical formula. They start with a default state. You pass it data, it will produce some output. However, this output is the default output of that algorithm. You then pass data over and over again to this mathematical formula. The formula now tweaks itself. We call this hyper-parameter tuning in the case of simple ML algorithms. In the case of a neural network, it is called adjusting the weights. But both approaches, effectively do the same thing. Keep tweaking the mathematical formula.
When you have passed enough data, to have successfully tweaked the mathematical formula, the formula would have changed enough to now be representative of your data. This is when we say that the AI system is now fully trained. If you now use this system, it will do what you trained it to do. Humans learn with repetitive learning. When we have said the same word to a newborn again and again, finally the newborn can speak that word. Pass enough data, and you will re-write the neurons of the human brain to be representative of the data you passed. Because AI systems learn the same way as humans do, they can learn to do most things that humans can do. This is where the name Artificial Intelligence comes from.
Now the last part. We have a single human brain design that can do everything. But we have all these various AI algorithms. Why is that the case? Why do we need like 100+ different AI algorithms?
If I tell you to learn a new word in the English language, it would take you like 2 mins to learn it correct? Understand the meaning, understand how it is pronounced, and understand how to use it in an actual sentence. 2 mins are all it would take you. What if I were to tell you to learn a new word in a different language say Spanish, Japanese, German? Let’s consider you don’t know these languages. You have never learnt any of them before. Now can you learn the new word in 2 mins? That is understand its meaning, understand how to pronounce it, and understand how to use it in sentences? Not quite possible in 2 mins now is it?
AI algorithms have essentially the same problem. If you take an AI algorithm that understands English and try to teach it a new word, it will learn the new word quite easily. However, if you take an English algorithm and try to teach it to speak German, you are in some trouble. It will take you an insane amount of data and time to get it trained. This is exactly the state of AI in the real world. We have all these different algorithms because their default state suits different types of AI needs.
If you want to build a recommendation engine, then you start with an AI algorithm that generally understands recommendation engines. If you want to predict the stock market, you start with an algorithm that works well on numerical time series data. So on and so forth.
The closer the algorithm is to your nature of data, the faster it will train. Faster training, also means you need lesser data to train it. However, this does not mean anything and everything is possible. You could get an English speaking algorithm to learn German, but you most certainly cannot get it to trade on the stock market.
If you keep speaking to your dog, you can train it to follow your command. It can understand what you are saying and will react to it. But no matter what you do, you cannot teach your dog to speak in English. That is not possible. The neural networks in the dog’s brain are only capable of a woof. No matter how much data you pass, and how much you re-wire them, those neurons can never speak like a human. They can only produce a woof. Most certainly in different tones if that is what you are aiming for.
AI works the same way. Not all AI models will work for every requirement. You must choose one that is relevant to your requirement.
I hope this article offers you a simplified understanding of AI. To summarise, AI is nothing but a bunch of mathematical formulas. All formulas have a default state. We train them to bring them closer to the state we want them to be. Training requires us to repeatedly pass data to the formulas until they have successfully tweaked themselves to be similar to our data. When the model has learnt enough, it becomes what we say artificially intelligent.