You made it to Module 6, well done, by now you've learned what machine learning is. The various phases within a project, and several considerations when it comes to your data. Including potential bias that can be amplified through your project. Now that you have a thorough understanding of the foundations. In this module, I'm going to help you exercise your creativity by focusing on ways you can discover ML use cases in your day-to-day business. I've organized the strategies for discovering ML use cases into five general themes. They are rule-based systems. I'll talk about what they are and how you can simplify or replace them with ML. Next, I'll cover how to automate business processes with ML. And understand unstructured data such as open text and audio with ML. Then I'll explain some ways to personalize end user experiences with ML, and explore creative uses of ML. All of these will build on your foundational knowledge. I'll then close the module with a two part exercise involving a worksheet and a hands on lab on sentiment analysis. Let's begin. [MUSIC] We often find three areas of day-to-day business that offer opportunities for using ML. They are replacing or simplifying rule-based systems, automating business processes, and finally understanding unstructured data. Let's look at each in turn. Most computer programs have at their core a set of heuristic rules or rules with procedures, that help users complete a task or make a decision. Suppose, for example, you manage a global retail company and you need computer equipment sent to a new branch in Amsterdam. You might have a rule-based system to help you choose the right supplier. Your rules for choosing the right supplier are that the company consistently delivers on time, within a specific budget. And with high customer satisfaction ratings. You then get a price quote from one global supplier, another from a European supplier, and a third one from a Dutch supplier. The rules would help you make a more informed decision, by determining which supplier is the preferred choice based on their last three order deliveries. But this was just for one decision. Inevitably, rule-based systems will either be limited in performance, hard to maintain, or limited in scope. Let's look at an example where ML can be used in a business context to simplify or replace rule-based systems. The scenario that I'll describe in a moment is based on a project that we at Google Cloud did with one of our customers. We've changed the company name and some of the details for this lesson, but the core ideas are the same. Will call the company Acme Widgets, which makes sophisticated smoke alarm systems for home use. When a customer buys a smoke alarm system for their home, Acme Widgets sends an agent to install the alarm. The agent's job is to first disconnect any existing system, then install the new one. Here's where an agent would have to use rule-based procedures to do their job successfully. There are many smoke alarm system manufacturers. So the agent needs to recognize the wiring for the existing system, and then interface it to the new system. Next is the sophistication of the new system. It not only measures smoke, but it also makes cost emergency teams in the event of a fire using a nearby phone system. So the agent would need to then integrate the new system with the phone line. The agent needs to know whether they can reuse the existing phone line or if they have to install a satellite connection. As you can tell by now, all of these decisions are quite complicated. Acme Widgets has a manual for its agents to help them make the right decisions at every point of the installation process. But the manual has hundreds of pages and thousands of schematic diagrams. Naturally, it's easy for the agents to make mistakes. They might look at the wrong page, miss step in the directions, and ultimately installed the system the wrong way. Now, how might this company use ML to replace or simplify its rule-based installation process? To start the company can list the rule-based decisions to agents, need to make in a sequential order. And replace or simplify the ones that are most challenging and with the most interdependencies. Based on some of the agents decisions I described earlier. Identify the make and model of the existing alarm is the biggest determinant of a successful installation. Next is the agents ability to follow the correct set of instructions to disconnect the existing alarm, and match the wire to the new system. What if Acme could create an ML model that would recognize a customer's existing setup? That is the manufacturer, the model, and the wiring. How it would work is that a pot arriving at the customers home, the agent would take a photograph of the current system. And then feed it into a custom trained model for the correct identification. That's not all. If the ML model was trained to identify the manufacturer, model and wiring of the existing system. Then it could simultaneously be trained to output the corresponding steps to disconnected, and the procedures for connecting the wiring to the new system. So now all the agent would have to do is to first take a photo, and then read the procedures off of their portable device. But wait, can ML do more to help an agent at this point? Think about it. What if it's not practical or efficient for the agent to read the steps? You can take the ML solution one step further by having it read the steps out loud. This is especially helpful in scenarios where the smoke alarm is difficult to reach. The agent can even ask to hear the instructions repeated, or to skip a few instructions. The possibilities for efficiency here are endless. Now let me ask you an important question. What data should Acme use here to train the ML model? We know it's images of existing smoke detectors. You might think that all they have to do is to photograph all of their existing tests units. Or to obtain photographs of all the smoke alarms referenced in the manual. And in this case, the company already had test units. So they could get a team to photograph the units, and train the model to identify the manufacturers and output the matching manual. While the model could be trained this way, it poses a problem. The test units are in pristine condition, while in customers homes there may be dust and cobwebs over the units. It's important then to train the model using data it's bound to encounter in the real world. Acmes agents don't want to take professional photographs. They want to take photographs with a mobile camera from any angle and with various lighting. So the company would have to ensure the data they used to train the ML model is either a simulation of dust and cobwebs, from various angles and in different lighting. Or they would start a data collection process first using actual in home systems. And then build them a model once they have enough data. So remember, wherever you're making rule-based decisions, there's an opportunity to use ML to replace or simplify your decision making process. In fact, the benefit of using ML is that you would scale better. And you can train the relevant ML models faster than the time it takes for humans to update rules in a computer program. Move on to the next video to learn about the other two ways to discover use cases for ML in everyday business.
