How is it possible for a computer to recognize a face? How does YouTube know you’re going to like the video that it recommended for you before you’ve watched it?
Artificial intelligence and machine learning have been in the spotlight for the last few years. They’re the buzzwords companies are putting into their public statements to boost investments and public interest. Headlines are written about the biggest tech companies utilizing it in all of the core functions of their businesses. But what most people don’t realize is that machine learning is built on the foundations of more universally recognizable statistics.
The rapid advancement of the field has been met with some trepidation, with big names like Elon Musk warning us that, if it’s taken too far, we could soon be at the mercy of our machine overlords. His perspective is that if we are to continue on this path now, we must start preparing for it through regulations so that we aren’t caught by surprise.
It might be comforting to label him as an alarmist, but it’s also irresponsible. Musk’s contribution to modern technology is undeniable; and therefore his warnings should be met with some legitimate consideration.
A Broken-Down Explanation
Although they are used interchangeably, artificial intelligence and machine learning are slightly different, with the latter being a subset of the former. As one could imagine, a machine making intelligent decisions based on a given criteria (AI) is the broader picture in which a machine that learns and makes its own decisions (ML) exists.
Broken down to its roots, machine learning is statistical analysis that’s been going to the gym for the last 60 or so years off and on, but has been really hitting it hard for the last 30 years. Where statistical analyses performed by humans tend to be laborious and time-consuming, computers have the unique ability to perform lots of simple tasks much more quickly than humans can. As research in more complex statistical methods has advanced, so too have computers. With the capability of performing standard statistical analyses in a fraction of a fraction of the time that a statistical expert could, further refining of these techniques in machines will only make their ability to understand statistics even greater.
At the risk of oversimplifying the emerging field of machine learning, that’s really all that machine learning is — hyper-advanced statistics. By using a backlog of data that a machine is able to quickly and readily parse through, a machine learning model can accurately predict future results due to nuances in the data that it has seen.
A Digestible Example
To give context to all of this jumble of words, I worked on a project where a machine learning system would identify authors based on their unique writing styles. I used Jane Austin, William Shakespeare, H.G. Wells, and Charles Dickens as my authors that would need to be differentiated. Then, after cleaning up the text so that each piece had a uniform structure, I extracted data from the text that I thought would be useful in differentiating each author. I used things like the percent of adjectives, verbs, and nouns in each sentence. The number of words per sentence and the number of sentences per paragraph and more to get a set of data that would hopefully give a unique quality to each author.
Initially, I used what is essentially a linear association between all of the data. You can imagine basically a graph that takes certain data, plots it on a line, and is able to make a rough prediction based on the information it has seen. It didn’t work very well. It was better than random guessing, but not by much.
I moved on and tested more and more methods until I used something more complex that acted as a decision tree that would make a ton of decisions based on a ton of combinations of data that would lead it to make very educated guesses based on which branch of the decision tree that a piece of text would follow. That was much more effective, achieving the right answer about 93% of the time.
This machine would more than likely outperform a casual reader in identifying these random pieces of texts from these authors. But showing them to an English expert might not be as impressive. But it would certainly be able to get through more text in a shorter period of time. Implementation with a focus on plagiarism detection would probably be an appropriate use-case for the tech because the authors wouldn’t be four of the most well known in existence.
If I, as a college student, was able to utilize machine learning technology like this in a way that can result in powerful consequences, just think of the implications of larger corporations using all of the data they are able to amass with all of the great minds and powerful technology that they are able to harness. And the technology is ever-evolving.
It is, however, important to understand that while this science is certainly difficult to grasp in its minutia, its foundation is something that everyone is familiar with. And despite how it may appear, machine learning doesn’t make decisions like humans do. Even the controversially named neural network doesn’t think like a human. They think like computers. They make lots of small, rapid calculations that lead them to a final decision.
Where humans have an innate understanding that when they see someone smile, they know that the person they are looking at is happy. Computers, on the other hand, have to see the curvature of the mouth, the eyes squinting. Then they need to compare that face to its log of data that it has on people’s expressions to see if the person is sad, confused, annoyed, irate, or happy. Perhaps the decision can be made in the same amount of time that it would take a human to process the emotion, but the process itself is fundamentally different.
This is part of the reason all of this is weird and scary. Machines are coming to the same conclusions that humans are coming to and even with an understanding of what’s happening under the hood, it’s just kind of disconcerting. And when you start to realize the incredible advancements that have occured with AI and machine learning in just the last five years, it’s hard not to be weary of what the next 20 years could produce.
Computers are taking people’s jobs now as well. Even tasks that are seemingly advanced are at the mercy of an educated computer. Machines are able to produce art that is convincing enough that people cannot discern it from manmade artwork.
Will computers be able to build on artwork just like humans have been able to do? Will computers be able to make scientific advancements independently? If the answer is yes then what will the landscape of jobs look like? Will there be any jobs?
Right now we simply don’t know; that’s how new this technology is. But there is the distinct possibility that this could change everything we know and it’s happening fast.
And to think that it’s all based on that stuff you learned in your stats class in high school is pretty crazy.