Supervised Learning #
Supervised learning involves a human “teacher” guiding a computer program (known as an algorithm) to capture patterns in historical data that contains known input-output mappings (“labeled data”) and create models that can be used to make predictions on new data, never-before-seen data.
Supervised learning is responsible for the overwhelming majority of the economic value created by machine learning today [Ng, Andrew at 13:29 minutes]. Examples of application are estimating the price of a car given its mileage, age, brand, engine capacity, etc, and predicting the likelihood of a prospect buying given their personal attributes, economic conditions, and features of the item being sold.
Unsupervised Learning #
Unsupervised learning involves an algorithm deducing the underlying structure of data that does NOT have input-output mappings (unlabeled data) by identifying similarities and anomalies in the data without a human teacher. Examples of usage are segmenting customers to better assign marketing campaigns and detecting outliers in banking transactions to identify fraud.
Semi-supervised Learning #
Semi-supervised learning is used to train models on partially labeled training data, usually a lot of unlabeled data and a little bit of labeled data. An unsupervised algorithm places the data into clusters then a supervised learning model built on the few labeled data points assigns labels to the unlabeled data in each cluster. Applications include text document classification and face recognition.
Reinforcement Learning #
Reinforcement learning is a behavioral model. The learning system (an “agent”) learns through trial and error, rather than training data, using a feedback loop of “rewards” and “penalties” to guide it towards the best outcome. The agent learns, over time, the best strategy (“policy”) that maximizes the rewards, thereby learning what action it should choose when it is in a given situation. In simple terms, the agent learns in the same way a dog learns. Applications are optimizing the trading strategy for an investment portfolio, balancing the load of electricity in varying demand cycles in a data center, finding the path that maximizes student outcomes in a course, etc.