I am a new Dad. Our son, Finn Theodore was welcomed into the world 14 weeks ago and it goes without saying, it has been an amazing experience. I draw a lot of inspiration and ideas from a diverse range of fields and whenever we are building products at CluedIn, we often lean on a lot of the ideas and inner workings of the human body to give us design ideas. It was this method of inspiration that made we realize the obvious connection between Machine Learning and bringing up a child. We are bombarded with quite a lot of scare mongering around Machine Learning and Artificial intelligence, but I would like to take a different stance – to embrace a different vision. Having a child is a grand responsibility and in here lies my thoughts on Machine Learning and Artificial Intelligence. We as Engineers, as individuals, as Data Scientists and as Companies have a responsibility that if we are building machines that are self learning, that analyze data and make decisions that are instantly actioned, then we need to be responsible for this. To reinforce my position, I want to take you through a slightly fictitious story. Hopefully along the way we can have a laugh about the situation and put some perspective on things.
This is my boy, Finn.
This is my other boy, HP Bladesystem c7000. We call him Henry.He is the brains behind CluedIn.
Finn just started to find his hands and is grasping at toys, at us, and is learning a lot more about the sounds we make and the sounds he makes. Henry is calculating complex math while combining 24 dimension matrices to calculate the statistical confidence for our merging engine.
It goes without saying, we are very proud of both their progress.
It also goes without saying that I spent a lot less time with my wife in conceiving Finn, than I did with Martin, our CTO, in conceiving Henry.
Yes, this is the weirdest sentence I have, and probably will ever write in my entire life. However it was only after we had Finn that I realized that the key to us building Henry was that we had a responsibility now to make sure that it was not doing anything that would be deemed as bad.
In a broad way, when we talk about Machine Learning, we typically break it up into Supervised Learning, Unsupervised Learning and Reinforcement Learning. Each individual technique can be quite easily associated to how we will bring up Finn in this world. Starting with Reinforcement Learning. Enter Psychology 101: Reinforcement and Reinforcement Schedules.
Like machines, a lot of bringing up a child is about teaching them that fire hurts and that being polite is good – translation: Positive Reinforcement, Negative Reinforcement, Punishment and Extinction. This is also exactly how Henry is taught as well. A simple example is that we guide him through what data is sensitive, what decisions he made are incorrect and what parts he did that he did a really good job on e.g. we can do this in CluedIn by marking content as sensitive.
Just like Finn, and yes he is a bit young now for this to be relevant, but it doesn’t stop my wife and I telling him that when he sleeps well and that when he has a big smile on his face that he is such a good boy and we shower him with love. However when he is being a bit “grizzly”, then we make sure we tell him that there is no need to be like “that” and that he has the support from us to help.
I always recall at CluedIn when we brought our first customer on in the food industry. This world was new to Henry and hence he did not do a great job at detecting people in text. In fact, he thought that “Spicy Chicken” and “Lamb Korma” were peoples names. Although Henry couldn’t see us laugh, we knew that we had to tell him that what he had done was not correct and once we had trained him that these were not names, we gave him another run at it. Every new time we corrected him and reran the process, his precision got better and better -enough to the point that he probably wouldn’t not make a similar mistake in the future, given similar input. Then again we can’t forget that like we can be trained, to some degree, we can also untrain people – given bad data. In fact, we are seeing this more and more today due to the over abundance of input that we all have to consume, from so many diverse mediums. It seems the only way to get valid results is to cross review this with experts. I draw a close experience that I will have with Finn, when we are first teaching him how to speak in English and Danish and how we will need to go through many iterations of input for him to start grasping things. So as you can see, this reinforcement learning is a post process i.e. it is up to us to validate or correct after a certain decision has been made.
I have always thought that it is important for us to teach our children how to learn, not what to learn. I thought this was the easiest analogy to make between bringing up Finn and how unsupervised learning works. Just like you instruct your child when they play Tetris “Ok, the better your score, the better you go. You should try and place the bricks where it makes sense”. Seeing that computer games was such a fundamental part of my childhood, I would love to bring Finn up with similar experiences and hence until his cognitive abilities have increased, I expect that I will watch Finn the first few months as he aimlessly moves those blocks directly on top of each other, ending the game in 15 seconds. However it is only a matter of cognitive growth and time before Finn is beating his old man at this game. For Henry, this is slightly different. For Henry, he is given the advantage of having a brain that can do a lot more processing, much quicker. Right now, his precision level will not universally be as good as Finn’s will be, but that is only a matter of time.
The “Spicy Chicken” example above is a good case of when Unsupervised Learning went wrong. It was simply a matter of not enough data input for Henry.
My point, is that unsupervised learning is all about giving Finn and Henry a huge amount of data. Some of it is correct, some of it is not. Statistically, I am hoping that I give both my children enough correct input that they can make their own decisions in life. For both of them, it is mostly the same process i.e. collect, organize, extrapolate, test, learn and then ask for more data.
My wife and I did a lot of preparation before Finn was in the world. We read books, we put little labels on the pages where it seemed like a good idea. We talked to the experts and we had great input on what we should do and what we should not do to bring up Finn. This is the same for Henry. We have seeded Henry with a plethora of labelled data where we have done the best we humanly can to label certain parts as incorrect and others as correct. Although this is susceptible to human error, we assume that Henry will also allow for this error rate. With all of this labelled data, the bottom line is, it will never be enough to stop Henry from making some really bad life choices.
So what does this part of the story look like for Finn? Just like we will try as hard as possible to tell Finn that for his first highschool party that he should not drink alcohol or smoke, it is only a supervised learning where he can get input from a reputable source (his parents), but then inevitably he will weigh up all the other pieces of data e.g. Will this increase my popularity? Will it make me find my first partner? – and only after this will he make a decision that he will personally learn from by doing it. This is where we bridge into Unsupervised Learning where once he realizes the problem of drinking too much alcohol, he will use this input to make further decisions in the future. It goes without saying, I am not looking forward to picking Finn or Henry up from their first drinking party.
Kids grow up. They start to go their own way and make their own decisions. We need to be conscious that we should expect the same from the machines. But like I will always help Finn in his life, give him comfort, maybe chip in a bit for his first car. I will also always be there for Henry. Replacing his harddrives, correcting him when he has done something unethical and celebrate him when he does amazing things. And this is the key behind this satirical but yet poignant story. Like we have certain levels of ethics and responsibility when we are building products today, we need to maintain this level of responsibility as we go into a new era of technology. Just like every parent should be responsible for their child, I think it is important for us all to be responsible for the machines we build.