What is machine learning?

- [Man 1] So what is machine learning in general? And give me some insights as to, you know, the reality of it. - [Man 2] If you think of about AI as you know, think about Venn diagram or a chart, right? You have AI, which is this anything that you produce, any algorithm that looks like, that is smart you can call it AI. - [Man 1] Okay. - [Man 2] Right. Now, people started by putting rule-based systems. You would encode all the rules that if you do this, I'm going to do that. - [Man 1] If this, then that. - [Man 2] Yeah. So that is also AI, you could say, right? But then people figured out that that's very cumbersome and it's very difficult. So if you use data to make the inference, right? If you are using data, that's called machine learning, right? So solving an AI problem, using data by learning from data, from experience, learning from experience, rather than somebody, a human expert coming in and encoding the rules. - [Man 1] Interesting. So to me in the data science world, we've done this forever. But we're generally trying to predict say sales numbers. - [Man 2] Right. - [Man 1] So, but we'll look at historical, you know, trends in all the variables you can think of, seasonality, products, whatever marketing spend, those kinds of things. Would you call that machine learning, doing a prediction on a number? - [Man 2] As long as you're using data, you would call that machine learning. - [Man 1] Really? Okay. - [Man 2] Yeah. You would call that machine learning. - [Man 1] Okay, so I'll put that on my resume. Okay, cool. (all laughing) Keep going. - [Man 2] Right. - [Man 1] I like this. - [Man 2] But let's say you did not use pass, but you know that, oh, when I put a deadline in my core sales. (man 1 laughing) The sales, you know, on that deadline day, you get the most sales, right? - [Man 1] Right. - [Man 2] You know that from experience, right? You just encode that all. - [Man 1] That's human learning not machine learning. - [Man 2] Exactly. - [Man 1] Yeah. - [Man 2] So in your algorithm, you say that on the last day I'm going to get 40% of the sales, right? You just encoded that. It's still you know, expert knowledge, but you did not learn it from getting 1000 course creators to come in and give that data or what were your sales numbers, right? - [Man 1] So then, so machine learning is literally just some set of logic. - [Man 2] Yeah. - [Man 1] Even if I write, if this then that using data. - [Man 2] No, it's usually automatically derived from the data. - [Man 1] Okay. - [Man 2] Right, you're not telling the machine that, oh, I want to encode this logic. You just dump the data and out comes the result. - [Man 1] Right. - [Man 2] And it could be using different factors in an intelligent way. You're not telling it explicitly that, oh, I want to weigh this factor this way or that, right? That's why you're... So think about it this way, that you have an objective function, you want to predict something, right? You know the ground truth, which means from your data, in your training data, you know what you want to predict and you know the actual prediction. - [Man 1] Right. You write something called a loss function, right? Which says, if your predictions are getting wrong then the loss would be high. And your objective is to reduce the loss. So you encode something, you write some kind of machine learning algorithm whose job is to reduce the loss on the training set, right? And in a perfect case, your loss on the training set would go to zero, right? - [Man 1] Yeah. - [Man 2] And it would also generalize, which means that if you get new data, which it had not seen the loss would also be similar to... - [Man 2] Now, so that's brilliant. And it's so interesting how I've never thought of like what we used to call in the analytics space data mining as machine learning, but it's very much that case of... - [Man 2] It is very much. - [Man 1] Yeah, so then, okay, so that's a general definition of machine learning. - [Man 2] Machine learning. - [Man 1] So give me down, like if you're talking to other machine learning experts or computer vision people, what do you consider machine, isn't that same thing or is there more to it than that? - [Man 2] No, it is about that, right? - [Man 1] Okay. - [Man 2] Whenever we use data and we are training, we are optimizing a system so that it learns from data and you're trying to minimize the lost function to... You know, you have a logical function - [Man 1] So then when new data comes in, whether that be an image or a video, meaning lots of images sequentially or sales numbers or you know scuffing levels or whatever the new data is that algorithm, that by algorithm we mean a set of logic that has some output is going to be accurate. We're trying to predict something, whether it's is this a dog or a cat or are we going to have a good sales month or a bad sales month. - [Man 2] Exactly. - [Man 1] It's all generally machine learning. - [Man 1] It's all generally machine learning. And your hope is that what you have learned from the training set, generalizes the test set to unseen data, right? - [Man 1] Yeah. - [Man 2] Because you can always create a machine learning algorithm. - [Man 1] That's perfect. - [Man 2] That's zero error on the training set. Right? But that doesn't mean that it has learned the underlying truth, right? - [Man 1] Right. - [Man 2] So that's the challenge. - [Man 1] So, okay. So then... - [Man 2] (indistinct). - [Man 1] Well, I've had supervised and unsupervised. What we're talking about here would be which one? - [Man 2] Supervised. - [Man 1] Supervised? - [Man 2] Yeah. - [Man 1] Because we gave it a training set. - [Man 1] We gave it a training set, we gave it the input and the expected output. - [Man 1] Okay, and I use that... - [Man 2] That's the training set. - [Man 1] So we're looking at it saying, hey, here's your answer make the algorithm, give me answers like this. - [Man 2] Right.