AI ladder, a framework for deploying AI in your enterprise. AI Fundamentals. The topics that are covered include first, an overview of the fundamentals of artificial intelligence, and secondly, an introduction to our narrative story for the remainder of this course involving a fictional company called Amsel Health. The Fundamentals of AI. Everything in artificial intelligence is centered in data. So ultimately, if you want to create, for example, machine learning models to either predict or explain, you're going to need raw data. You're going to need raw data that is then processed into data that can be used to train machine learning algorithms. Those machine learning algorithms can then be use to make predictions for your enterprise, or to help you search and sort through data to explain certain phenomenon. That result is going to be based on the interpretation that you get relative to your business context. The bottom line is, you need data. Without good data, you will not have good predictions or good explanations generated from your machine learning models. One of the really tough things that some people have trouble wrapping their minds around is the fact that the results that you get from AI, and in particular for machine learning models, are that the results are probabilistic, not deterministic. What does this mean? Well, a deterministic result is something that leads from a logical conclusion. So in other words, if x equals y, then z. That is a very definite statement, with a very determinate outcome of calculations based on the instructions that are given to the machine. However, machine learning models do not return these definite types of decisions. Instead, they return what are called probabilistic feedback. For example, if you were to set up a machine learning system to make a prediction of sales based on some data set, the prediction will come back to you, but it will come back to you with a probability assigned to it. You will have this confidence probability assigned that can be anywhere from close to 0 to close to 100. It'll never be 100 percent. The nature of machine learning models doesn't allow them to be perfectly, absolutely right and deterministic all the time. There's always going to be a little bit of leeway when it comes to assigning a probability, and it will never be exactly 100 percent. But the higher the probability, the more the machine learning model is certain of its decision, or put another way, the more the machine learning model is certain that the inputs it received are similar to the inputs that it was initially trained on. So a machine learning model simply give you a probabilistic assessment of the outcome, which means it might be wrong sometimes. So ask yourself this question, "Can I use a data set of sales figures, sales incentive plans, and salesperson profiles to predict future sales?" That is a very standard and common-like use case for machine learning. You would start off with raw data sources. For example, employee profiles stored in HR databases, incentive information as to how, for example, your salespeople are being paid. Then finally, past sales data, past sales data that reflects rows and rows of columns laying out how each salesperson did relative to their incentive plan, to their geography, to their profile, to whatever else you happen to have data for. You can never use data as is. The data always have to be cleaned and sanitized. Typically, they have to be re-transformed and analyzed, looking for certain features that can be used by the machine learning model to make predictions or explanations. That process of cleaning up the data and bringing it in from its original sources can take awhile. After the data have been cleaned up, they are then used to train a machine learning model. The machine learning model is set to a training mode, where it is given data that has the correct answers and correct predictions as determined by the data scientist. Once the machine learning model has been trained on those data, the machine learning model is then switched on to prediction mode, then it is given data that is never seen before. Using its training, it takes the information that its never seen before and makes a prediction based on its training. So in this example, we provide the machine-learning model with a whole bunch of information related to sales and sales people and incentives and their profiles, and then we ask it a question about a specific salesperson, that is, "Ana is on Incentive plan 2 and located in the mid-Atlantic. What are her projected sales for July?" The machine will take this information, use it's training to come up with a prediction of her projected sales for that particular month. We just explored the foundations of machine learning. All instances of machine learning algorithm can do really only one of two things. They can either predict, or they can be used to explain why data is laid out and organized the way it is. You can use machine learning to predict some future state, given some present state of data, or you can use it to go through vast data sets and look for patterns that you can use to explain why those data sets are the way they are. Prediction and explanation both rely on multiple different types of algorithms, and those algorithms are fed by data.
