When you hear someone talk about Machine Learning or read about it in an article online. Oftentimes, ML is used interchangeably with Artificial Intelligence and sometimes Deep Learning. It's important to understand though, that these are different terms and they mean different things. I'll explain each one. Artificial intelligence, or AI, refers to machines that are capable of acting autonomously, essentially machines that think. There are many ways to make machines think, just as there are many ways that humans think. Machine learning is one branch of AI and one way to get a computer to solve a problem autonomously. If Artificial Intelligence or AI, is a field, Machine Learning refers to a class of computer algorithms that solve AI problems. Deep Learning is a subset of Machine Learning algorithms and refers to a specific class of Machine Learning algorithms called neural networks. Neural networks are used in many business settings today because they generally outperformed other Machine Learning algorithms for tasks such as image classification and natural language processing, and there's more. Deep Learning is a type of Machine Learning that works even when the data is unstructured. Images, speech, video, and natural language text are all examples of unstructured data; they don't exist in a tabular format. By contrast, structured data would exist in a tabular format, like credit card numbers, gender, or age. It's important to note that not every Machine Learning algorithm is Deep Learning, but every Deep Learning algorithm is Machine Learning. Let me give you a couple of examples to illustrate this. A decision tree is a Machine Learning example however, it's not Deep Learning. RankBrain is a neural network used by Google search and it is a Deep Learning example. For the purposes of this course, let me give you a formal definition of Machine Learning that will guide your learning. ML is a way to use standard algorithms to analyze data in order to derive predictive insights and make repeated decisions. That may sound complicated, but all the parts of that definition are important. Think of it like the recipe for ML. I'll go over each one of these ingredients in more detail later in this course. First, let me clarify what I mean by predictive insights with a few examples. Searches, of course are our flagship application here at Google. It wasn't able to make predictions based on user searches. However, over time, our engineers used Machine Learning to improve the search experience. This is a typical Machine Learning use case, I'll break it down. Suppose for instance, you want to search for The Giants, a sports team. But wait, if you type in giants, should the search results show you San Francisco Giants or New York Giants. One is a baseball team based in California and the other is an American football team based in New York. How would you program the machine to give users the correct results? A few years ago, this is how Google search worked. The search engines' code-based used rules to decide which sports team to show a user. If the query is giants and the user's in the big area show them results about San Francisco Giants. If the user is in the New York area, show them results about New York Giants. If there or anywhere else, show them results about tall people, and this is just for one query. If you multiply this by thousands of different queries and by different users each day, you can probably imagine how complex the whole code-base had become. The reality is, hand-coded rules are hard to maintain. Why not try Machine Learning? You might also remember that one important part of the ML definition is that it works well at scale. ML scales better because it's all automated. If we add all of the data that tells us which search results users clicked on per query. Why not train a Machine Learning model to predict the rank for the search results? That was the essential idea behind RankBrain, our deep neural network for search ranking, which was introduced in 2015 by Google's engineers. It outperformed many human-built signals, and using ML we were able to replace many of the hand-coded rules. The neural network ended up improving our search quality dramatically. In fact, Google's neural network is a key differentiator among similar technologies in the market. An added benefit of RankBrain or any machine learning model is that the system could continuously improve itself based on new user queries and new user clicks. That was an example of using machine learning to automatically predict the rank of internet wide search results. Another great use for machine learning is for automating processes. These types of business problems highlight the benefit of using ML to make predictions and repeated decisions at scale. Let me give you an example. Ananda Development is a property developer headquartered in Thailand. One of their most important business processes is handing over a completed condominium to a buyer. Before the buyer pays for the condo though, an Ananda Development inspector and the buyer conduct detailed checks of the condominium for any build variations that must be fixed. Ananda Development inspectors would visually check for problems and list any issues on paper and a prospective buyer might also take notes and photographs of the findings. On average, a single inspector would have to check several 100 items per day. Multiplied across several inspectors and multiple projects, this Workload adds up. Furthermore, this laborious manual process will subject to occasional human error. That meant data could be omitted or recorded incorrectly. So the business decided to develop a mobile application that would help inspectors identify, capture, and submit defects while they carried out their inspections. Ananda Development used machine learning to make it the inspection process more efficient. The app needed to recognize and convert Thai language speech and a version of English spoken by many Thai people to text. The company used Google speech to text API, finding that the product had an accuracy rate of 90 percent in recognizing Thai speech, and high accuracy rates in recognizing Thai English. The inspection process is now more lean and inspection times have reduced with more accurate outcomes. As another benefit, buyers also receive copies of electronic inspection reports and updated status notes as defects are repaired. Plus, inspectors can now complete their activities during the workday instead of spending extra time at the office filling out reports. As with most ML projects that streamline a business process, another development benefited by doing these: capturing operational insights and converting them into valuable actual data, delivering a centralized content management platform that resolved the challenges of communicating with multiple groups, providing real-time onsite monitoring that enables managers to make better decisions, analyzing construction and vendor quality to deliver continuous improvement, making statutory disclosures required for compliance and safety checks. They were also able to proactively initiate defect detection. That's a lot of added benefits for automating one business process. Let's look at one more example. Another class of ML use cases is for understanding unstructured data like images, videos, and audio. Video monitoring systems have existed for decades, but until recently, the videos were used for surveillance purposes only. In 2017, Arcules was formed to mine data compiled from video cameras and internet of things devices for clients who want to insights to improve safety, operational efficiency, and top line growth. For example, for retailers Arcules platform enables insights for retail footpath, sentiment and dwell time analyses. They can use video cameras and IoT devices to correlate data on shopper behavior in the store. Using video cameras and motion sensors, they answer questions like, did the shopper pick up the item? What was their sentiment? Did they consult their phone? What other items did they pick up? Did they make a purchase? With analysis of this type of information, the store can make adjustments to influence shopper behavior. It's important to know that online retailers already have all of this information. They know which pages you browse, which products you'd look at next and whether you add the item to the cart. Systems like the one that Arcules builds enabled brick and mortar retailers to get the insights that online retailers already have by taking advantage of the data that already existed in their business. As Eric Schmidt, Google's former Executive Chairman pointed out a few years ago, the programming paradigm is changing. Instead of programming a computer, you teach a computer to learn something and it does what you want. That is the power of machine learning and the next transformation. In fact, this insight is what transformed many of Google's own products. In the next video, I'll talk about machine learning projects and how you can start working on them.
