Machine Learning is a discipline inside of artificial intelligence. The term AI was originally coined in 1956, and has since evolved to encompass many fields of study that are commonplace in today's technological discussions. Topics like Machine Learning, Natural Language Processing or NLP and Computer Vision, all fall into the modern umbrella that is AI. Now, if you're thinking of AI is killer robots and take over the world, your field of interest is more each HLI or Human-Like Intelligence. You'll notice that deep learning is yet an additional sub-discipline within machine learning and it's captured a lot of attention as it has started even arrival and surpass the human ability to perform complex tasks like image recognition, speech recognition, language translation and much, much more. Now that we know when AI, ML and deep learning came about, how about a one sentence definition of ML? ML and its core labels things for you. Show a model a bunch of good historical sales trends data for your clothing store and a model can predict next month's sales. Show model lots of photos of cars and what the correct make a model is, and with enough examples, the model will classify new unlabeled car photos for you. Let's dive a bit more with an example. On the weekends, you'll likely find me scouring the web for new sci-fi movies and TV series to watch. Now, I can tell you what previous sci-fi movies I like and I can do a decent job at narrowing down the list of potential new options myself, but I don't have all the time the world is scan through and classify good sci-fi movies to watch. I do have a general intuition of what I like, which is also reflected in my history of movies watched. So, things are has been set in space. In the near future, It's shorter than two hours and there's no crazy aliens and no horror or anything like that. I can provide you the list of movies that I also liked. We can then train a model that label in this case, it's classifying, whether I like a new sci-fi movie or a series that comes out which is about space drama. The key difference though, is that although I have an intuition of what I like and I'm providing you the model with the list of movies that I liked and didn't like, I'm not providing the model with a hard-coded recipe for narrowing down the movies. If it's shorter than two hours then prioritize this space movie as long as it's not horror, that sort of thing. The beauty of ML is that it comes up with this recipe by itself based on the correctly labeled examples that it has seen so far. Now, imagine if I didn't provide any insight behind my movie selection process other than all the movies and the TV series in the past. Would you have any basis to even build those hard-coded rules? Not anymore. What if I asked you to predict across all genres of movies, which can have very different aspects, how could you maintain a rules-based things like, if comedy and actor equals John, else if not horror and duration less than two hours, all that gets unwieldy. Let the Machine Learning model figure out the recipe that ties your historical labeled data to the predictions on unseen data. Now, let's extrapolate this to a real-world application. Let's take Google Search for example. Say you go to Google and you search for giants. What should we show you as your results page to make the most relevant for you. Or if you're in California like me should we show you the results for the San Francisco Giants, it's a baseball team maybe list some local games nearby. What about if you're based in New York? Should we tailor the results to show the New York Giants football team instead as one of the rules? Well up until a few years ago, this is exactly how Google search worked. There were a ton of rules that were part of the search engine code-based to decide which sports team to show and where based on where the user was. If the query is giants, and the users in the Bay Area, show them the results by the San Francisco Giants. If the user is in the New York area, show them the results but the New York Giants. If they're anywhere else, show them the results about tall people, giants. Those of you have worked with SQL before, just imagine how many case statements this would be and how hard it would be to maintain, and that's just for one query. Multiply this by the large variety of queries that people make, where they make them from, what the vices that are on. You can imagine how complex the whole code base had become. The code base was getting really unwieldy. Hard-coded rules are hard to maintain. This is exactly where ML comes into play. It scales much better because it requires no hand-coded rules and it's all automated. Our data set in this case, is we know that historically, in the search engine result pages which links that people clicked on. Why couldn't we just train an ML model to provide input into the search ranking? That's exactly what Google itself is done internally, and they used a deep learning ML model called Rank Brain. After rolling it out, the quality of search engine results improve dramatically with the signal coming from Rank Brain becoming one of the top three influencers for how results are ranked. If you're interested, I'll provide a link where you can read more about it. Now to a recap. In Machine Learning, we want to lead with examples, not with instructions. Even in business applications, we have these long case statements of; if, then, else and logically hard-coded all that stuff together. But you do have a history of good labelled data. That's a possible application for Machine Learning. Now deep learning. Remember that's a sub-discipline of Machine Learning is useful for we as humans can't even map out our own intuition about what makes a prediction correct or not. So, what do you see here? Now your eyes and your brain and the benefit of many years of evolution and intuition to allow you to perceive and interpret all those pixels on the screen. How can we teach a machine to understand that this picture here is a cat? You let you fall yourself back into the rules making bad habits that we're trying to avoid, you might say well look for a cat-like eyes and these images. Okay. What about this image? Your brain still knows it's a cat, but the machine now has no basis to go off even of our old rule of just look at the eyes and determine if it's cat-like eyes. Okay. What happens if we added a bunch more hard-coded rules like this; look for the ears, the eyes and the nose? All right, is this still a cat? What about this? Again you get the point, hard-coded rules completely fails us here, and that's where deep learning comes into play. Before, we just have labeled examples and we completely let the model figure out how to build a good recipe to answer the question, what is a cat? In 2012, that's exactly what the Google Research team with Jeff Dean and Andrew Ng did. What you see here is what the deep learning neural network figured out, what a cat is based on looking at over 10 million images and processing the model over 16,000 computers. Now, a familiar architecture for deep learning is the neural network, which is the model inspired by our own human brains. Here, it takes the input image that you see there and classifies it as a cat or a dog. Again, we're not telling the model to focus on looking for dog collars or a cat whiskers. It builds its own recipe for determining the correct label and applies it to the end. As you can see from the image, modern ML models can scale and handle even tricky data points like this dog hiding in the laundry basket.
