Types of Machine Learning
There are so many buzzwords and terminologies that have hit the mainstream by now that approaching the world of Machine Learning and its Jargon can be a bit overwhelming. This post will define what types of machine learning actually are, and give a few examples to help reinforce those assertions.
What is a type of machine learning?
It’s basically the method you use to bring a model and its outputs into the state where they are producing the types of results that are expected.
To take that a little further, let’s say you must create some process like this:
An environment is allocated where some set of data can be provided
A pipeline for feedback can be added
Some relationships between the system's inputs and the model are established
The outputs of the system are collected and measured
A type of machine learning describes the approach a specified algorithm uses to learn from its inputs, the nature of the feedback (if any) it uses for learning, and the kind of task it’s meant to carry out.
Let’s dig into this a bit more by looking at three basic types of Machine Learning.
Supervised Learning
Supervised learning is a type of machine learning that has a set of training data that is labeled (traditionally by humans but also by other models) and assumed to be accurate.
This allows the model to categorize, or assign scores and weights to outputs. After the model is trained it can accept unlabeled data and provide its predictions based on correlations it builds based on the training data.
As a “type” of learning we follow the rules above and say:
Its inputs during training are labeled data for classification or scoring
Its inputs after training are sets of real-world data
The feedback it would receive would be updated training data with new labels
The kind of task it’s meant for is to be repeatably accurate in its very specific predictions
Unsupervised Learning
Unsupervised Learning is a type of machine learning that is set up to identify “insights” within or between unlabeled data sets that have not been previously identified.
Its primary goal is to understand the structure or distribution of the data, find patterns, or transform the data in ways that make it easier to understand or use in subsequent tasks. So it’s automating the work of digging around in data sets that might be very large, or even very difficult for humans to understand and create items that we mere mortals might deem to be an insight.
As a “type” of learning we follow the rules above and say:
Its inputs during training are not labeled
Its inputs after training are sets of real-world data
It may not receive feedback but its feedback would be meant to tune the model so that it discovers more or less patterns or anomalies
The kind of task it’s meant for is to identify correlations, commonalities, structures, and even anomalies within data sets that aren’t labeled.
Reinforcement Learning
Reinforcement Learning is a type of machine learning where an agent learns to make decisions by taking actions in an environment to maximize some form of “reward” or “penalty”. The agent uncovers optimal or near-optimal sequences of actions in situations not explicitly pre-defined based on the amount and frequency of the reward.
In other words, we create some values that the system observes, and based on the magnitude of those values it can determine whether or not it’s prediction or action was appropriate. The better the feedback the better the model can be tuned.
As a “type” of learning we follow the rules above and say:
No initial training data is required
Its feedback is a value for reward, and possibly a value for a penalty to describe to it how well it’s done its job
the kind of task can be very widely defined or complicated, as reinforcement learning can even be used in physical motion systems.
Conclusion
The types of machine learning do not describe any specific model or formula, or any other very technical thing. They basically describe the job the model is meant to do and the way the model comes to understand how to create the desired outputs.
