Videos - The Coin Bot

Jump to This Time	Transcription Text
00:00:00	Machines are much, much dumber than the human that created them. What's up, everyone? Welcome back to the channel. In our last video, we talked about why this channel is called The Coinbot. We also talked about my interest in coin collecting and my desire to create a machine that's capable
00:00:20	of finding specific coins to add to my collection. I'm sure most of you have seen coin sorting machines in your local grocery store or your bank, and you know that they're capable of counting out change pretty quickly. Well, the challenge here is that we're not actually just trying to sort out coins by size, which can be done mechanically.
00:00:40	We're actually interested in finding specific designs, specific errors that are numismatically interesting, and that I'd want to add into my coin collection. If we were interested in only finding the pennies in a jar of coins, that would be pretty easy because we could do that by size. But in this case, we want to look at specific details, which makes it much harder and requires artificial intelligence to do that.
00:01:09	In this video, we're going to be talking about all of the challenges, or at least some of the challenges of creating that machine and why that is so difficult. But first, I wanted to talk about the differences between the way a human thinks and the way a machine thinks to kind of shed some light on those challenges.
00:01:27	First of all, a human is very good at general intelligence. That means they're good at using the context of their environment and also their experience to fill in gaps in their knowledge. A machine is not very good at general intelligence. In fact, machines right now in their current "state of the art" are only good at performing specific tasks.
00:01:51	And I'm sure you've heard it said that computers are only as smart as the person who programmed it. Well, that's actually not true. Machines are much, much dumber than the human that created them, and that presents a huge challenge. When sorting coins, for example, it's very easy for a human to pick up
00:02:11	a coin and easily recognize that as the 2021 D cent from the U.S. Mint. But for a computer, they need to be trained to recognize those. The challenge for computers is that they don't have that context or experience, so they must be trained to perform specific tasks that we want them to do. In our case, the computer is challenged by many things, even when we try to train it properly.
00:02:44	That brings us to our first challenge. The U.S. Mint has created over 14 different designs for just the U.S. Small sense alone. That started in 1857 with the Flying Eagle Obverse and is continued through today with the Lincoln Obverse. Some of those designs are shown here. The next challenge we face is the orientation of the coin.
00:03:14	If you train your AI to think that this is a Lincoln Obverse and this is a Lincoln Obverse and you continue to do that, showing them only coins that are oriented in this way. When the AI is presented with a coin that's oriented differently like this, It will become completely confused. So on top of those 14 different designs,
00:03:36	the orientation looks like a completely different design to the AI. So you need to account for that when you train your computer to recognize coins. Similar to orientation, the next challenge faced by the AI is changes in lighting. And that is because of the way that us coins are struck. They're struck in such a way
00:04:03	that the obverse and reverse designs actually are embossed. They create a 3D effect. And when you have different lighting effects, that highlights different areas of the coin. And again, if you've only trained your AI to recognize lighting from one direction, when faced with a coin,
00:04:21	with lighting from a different direction the AI gets confused because it no longer recognizes the highlights and shadows that are on the coin. One of the other challenges that I face is that I don't have a lot of experience with the software that I intend to use in order to design this machine. I intend to use Python, which is a programming language, TensorFlow, which is a machine learning algorithm created by Google, and something
00:04:51	called OpenCV to manipulate the images that we take of the coins. And in my limited experience with these applications, I run into another challenge that the AI faces and that's overfitting. That means that the machine is overconfident in what it thinks is a Lincoln obverse, for example, so if you are sending it multiple images and you send it this image and this image and this image, It might recognize all
00:05:22	of those coins correctly as the Lincoln obverse. But then if you send it this image, which is the wheat reverse. Then it would be confused by that. And it is 100% certain that that is a Lincoln obverse, when in fact, not at all. So I hope that helps to explain some of the challenges we face as we try to design a machine that's capable
00:05:49	of finding specific coins that I'd like to add to my collection. If you like this type of content, I hope you'll press the like button for this video. That you hit the notification Bell so that you don't miss out on any future videos and also that you subscribe to the channel as that would really help me out. In our next video,
00:06:03	we're going to be talking about some machines that I'm aware of that does something similar as well as some of the progress I've made in my own programming. Until next time, I hope you have a great day and I look forward to seeing you in the next video.