AI, Machine Learning and their limits
I know it sounds rather weird to start a conversation with limitations in mind, but, in my experience, this is best way to gauge the potential of a technology.
Now the main topic will be Machine Learning, which is the basis of how most AI projects are conducted these days. It doesn't mean that there aren't others (in fact they are and they seem quite promising), but it's what most companies in this sector are doing these days.
Wait a sec. What is Machine Learning more exactly?
It can sound awfully complicated, but you can understand its basics in less than 30 minutes.
Machine Learning is based on a computing structure called a neural network. It's not exactly intelligent, it's just able to take a variety of inputs and create some rules based on it using multiple layers. Here's a video explaining how it actually works. Trust me, after these 19 minutes you will know what ML is really about.
Back? Moving on, so what are the problems with ML?
Past results do not guarantee future earnings
It's what you hear whenever you invest your money, but in a sense this is the most fundamental limitation of Machine Learning. If adaptability is what defines intelligence, ML is definitely not the way to go. In essence, a machine learning algorithm takes past data and tries to apply to the future. Call it a training set, but it is what machine learning is really all about.
This means that in a context where the situation is prone to change rapidly and randomly, the technology is... well... relatively useless.
This is the reason for which ML is typically applied in contexts that are easily replicated and, if you can find such a situation, this is what you need. Games are particularly well suited for this type of work. The rules never change, so past lessons will always be valuable.
Take chess for example. Recently, Google's (or DeepMind's to be more exact) AlphaZero crushed what was considered to be the world's best chess engine and did it in such a manner that could only be called beautiful. It didn't feel "artificial" and the brilliance was hard to miss. IM Danny Rensch explains it better:
Garbage in, garbage out
Since ML is so reliant on training sets, the quality of the training set that it uses matters a great deal.
This becomes particularly important when you try to use data from "the wild". More often than not, what you will get is a subset that is not representative of the problem as a whole. If it gets really bad, some actors will try to actively influence the results.
Take this example: Microsoft tried to build a Twitter bot using ML. That was a bad idea. Tay, as it was called, quickly became the world's first racist bot, simply because the community over at 4chan wanted to have some fun. Giggles were had, lessons were learned, bots were shut down.
In a business context, this obviously has a bit more weight. How sure are you that the data you have is accurate? Answering this question may or may not recommend using ML in your specific case.
You're not the only player in town
Strictly speaking, whenever you attempt to use a technology, you need to do a cost-benefits analysis and the truth is ML is becoming a consumer grade technology.
Recently, a software was released that uses ML to replace the face of someone in some video with the face of somebody else. To the shocking surprise of absolutely noone, this is is now mostly used for porn. For the whole story check this BBC article.
Point is, the underlying technology is not that complicated and with the explosion in its use you will likely not see the benefits you expect. This isn't a race to the top, it is a race to the average.
Conclusion
With this in mind there are things you can do. ML is here to stay, but you have to be smart on how you use it.
Don't be afraid to take new paths, just because the data model doesn't have information on it. Protect your data as best you can and, most importantly, work hard to improve the quality of the data you get.
