We talked about human decision making and rule based systems were a software developer can put someone's expertise into software. Machine learning changes this paradigm a little bit. Machine learning is essentially pattern recognition. Machine learning does it uses examples. We call this training data to generate a decision framework. The key is that machine learning is really all about showing, not telling. Instead of a person explaining how they do something, what you do is you provide lots of examples, and the machine itself, learns how the decision framework is put together. This completely changes the workflow. It starts with the data not codifying the expertise. This is the key to machine learning, instead of providing a set of rules, instead of providing guidelines based on expertise, you just provide lots of data that serve as examples to the machine learning algorithm and the computer arrives in an optimal mapping. For example, instead of providing a lot of information about what factors or what variables, let's say we have a lot of information about candidates that are applying to the firm instead of providing expertise or providing a lot of information about what a particular organization values in candidates and how to think about evaluating different factors or prioritizing different candidates. What you might do instead with a machine learning system or in building a machine learning system is provide a tremendous amount of data on historical data, on applicants and these different factors to different pieces of the applicant portfolio and the final decision that was made about the candidate. Imagine you have this data. We have lots of information about the candidates and then you have information about what decision was finally made on the candidate, whether they were brought in for an interview or not. The second question, and we have this data, the machine learning algorithm can learn just from looking at the data, how they should think about or how it should think about weighting the different pieces of the applicant portfolio, whether it should prioritize prior employer experience, certain skill sets or so on. You never had to tell it. A human expert never had to tell it. Exactly what matters for that particular organization or context from the data, from being shown examples, it can learn what to do. This is the difference a key difference between the machine learning and some of the systems are some of the approaches that have been used before. This is a transformative development in a number of different ways. Why is that the case? Why is it that we think about the use of examples this way as being transformative? Why don't we think about machine learning systems being transformative? Number one is it changes the workflow, in the example of a rule based system, which we talked about. We need access to an expert. We need human expertise at some point in the process and that can be expensive, especially with a really specialized, highly trained experts. It can be difficult to get that expertise. Again, it can be difficult to ask experts like that to explain everything they know about a body of knowledge or an application and the machine learning case, I don't need an expert. I need the data generated by the experts. Maybe I can walk over to in a medical case, I can walk over to a hospital and get data on prior decisions or prior medical predictions made by practitioners. If I have that, I no longer need access to the actual expert. It's also transformative because there are some cases where you have so much data that there's really no way a human expert can do a robust job at synthesizing all of that information together. The examples we've been talking about, we've been talking about a few different pieces of a portfolio, but there are many machine learning cases or many information cases, information, context, where you have not three or four variables, but thousands of variables that are forming a prediction in the advertising space, in the finance space, sometimes in the environmental space, for instance, you may have thousands of variables. The notion that a human expert can really consider all of that information in one shot or give it a prediction, it's a little ambitious, but a machine learning algorithm can do a pretty good job. If you give it two thousand variables and you have those examples of what the final outcome was. The algorithm in this case can do a pretty good job and often a much better job than a human can who might just find it very, very difficult or impossible to synthesize all of that information at once. It's also transformative because of scale. Obviously, once we have the algorithm built, we can then run it at scale. We can. It doesn't matter if it's looking at 10 resumes, one hundred resumes, one hundred thousand resumes or 10 million resumes, it's pretty invariant to that. Once we have this up and running, it scales very well. It's also very consistent, which is a nice feature of systems like this, because it is an algorithm. It tends to arrive at the same decisions, given the same information. It doesn't get tired throughout the day. It doesn't start to make mistakes on some days relative to others. It's very consistent. In some contexts this really matters. Now, the key to all of this, the key to being able to build these systems and to achieve these benefits is all in the training data, these examples that you feed the system to build it. That's what we're going to talk about in the next video.
