We’ve Created Machines That Learn By Themselves…Now What?
I’ve been a fan of the concept of Artificial Intelligence for longer than I can remember, and I have spent the last four years of my life researching the human brain from an engineering perspective to try and understand how intelligence works (I’m not going to define intelligence here, I’m not falling through that rabbit hole), and more specifically, how we humans learn new skills. And I’ve been doing it with the hopes of eventually being able to get down in front of a computer and actually create an intelligent thing.
There’s this beautiful dance that happens between brain structures whenever we’re exposed to new stimulus, one part makes sure that you’re aware that you’ve learned this new skill, there’s another part that makes sure that the new skill you just learned can be performed without you having to consciously think about it (ex. walking: we just start walking, we don’t need to be thinking about each muscle and how to move it), there are other parts that make sure you use the right skill in the right moment, and even one that monitors the skill so you don’t mess it up. As I said, A set of brain structures that do exactly what you need them to do at the exact moment, just like a dance. From now on, I will refer to this as “The beautiful dance”.
What this beautiful dance allows humans to learn whatever skill we want to, the beautiful dance is what makes us general. Right now Artificial Intelligence is getting really good at learning specific skills, we have systems that can identify images, understand how humans talk, we have a medical assistant that is able to correctly diagnose a patient and the one that I find the most impressive, we have an AI that can compose music. However we haven’t been able to create a single artificial intelligence that can do those five things, and that my friends, that’s The Dream.
What do humans have that has been missing from AI’s? Well for starters it was hard for new AIs to stand in the shoulders of the ones that came before them (yes, I’m referring to AIs as thems, but that’s a discussion for a different time (over some beers obviously)), this was because of different factors but the most notable one, was the fact that AI has been built by companies, and most of what they do, they do for themselves. We didn’t had access to their datasets, we didn’t had access to their models, so we had to try and con our way to a good dataset to train our models (I actually had to do this to train the system I proposed for my masters), or the other option which is starting to collect our own data. Once we had the data, it was time to go to the academic literature and scrap whatever we could to create our models.
Fortunately that problem has been receding since Google and other companies started releasing their models and systems as open source libraries, now there are startups like Wit ai that give you free access to a really good natural language processor, IBM gives free access to Watson, so now any person can create an AI that solves a specific skill, and that’s great. But there’s this one thing that we’re still missing, remember the beautiful dance? How we had one part that took care of performing the skill? That’s what we have now, but we’re missing the other three parts, we’re missing the dance, and even worse, we’ve been ignoring it. So we need to start thinking about it, because without the beautiful dance, we’ll never have The Dream.
Originally published at medium.com on October 1, 2016 by Ory Medina. Ory is a Software Engineer at Wizeline and a PhD candidate at Cinvestav Unidad Guadalajara. His areas of interest include all things artificial intelligence and machine learning.