Modern machine learning is arguably the intermediate step between heuristics and artificial intelligence (AI). While AI will be a form of machine learning, the current generation of ideas we use will probably no longer be relevant when a breakthrough happens which allows for true artificial intelligence. Machine learning transcends traditional algorithms and rules when done right because it adapts to patterns which are harder to quantify and qualify. You don’t just map out rules, you show the system what you want, have, or let it experience the system and it seemingly learns from it.
Modern day machine learning isn’t magic (despite what SaaS companies may say), and ironically it isn’t even all that intelligent. It’s basically fancy math to turn a system into a process. You can either react to the inputs ( unsupervised learning), the inputs and the outputs ( supervised learning), or the environment ( reinforcement learning) with the primary paradigms. As these are neither inclusive nor exclusive, they can be combined to get more complicated systems and processes.
The model can affect how effective the learning is, but the paradigm is going to affect the process. A specific task is different than trying to make sense of data. An environment is different than fixed inputs or outputs. Let’s go over what each paradigm is and what it means for machine learning.
Supervised learning is a form of machine learning which effectively works as concept mapping. You have an input, and you get an output. As you feed in data and assess it, you get a function which tries to abstract the system in order to have rules that probably make no sense to an actual person. An image contains a car and your expected output is a yes, this image doesn’t contain a car, so you get a no. As the system learns, it can rule out non-cars with increasing accuracy.
When you teach certain concepts, you can’t really directly explain them and expect your audience to understand. You describe what makes that concept work but also give examples. There is a level of inference that if you have this input, you get this specific output, and they draw their own conclusions about what connects them. If you’re teaching a kid what a dog is, you don’t explain the legs, the eyes, the ears, etc., you point and say: “That’s a dog,” or: “That’s not a dog,” until they get it. You’re providing a task to be done and examples of the right and wrong answer.
For our car example, this would mean we’ve codified images as containing a car or not, and the system abstracts the pattern. The main limitation is that it relies entirely on the training data. If all your pictures are red cars in the woods for the car set, a tomato with green around it might be determined to be a car because there is a red shape and green around it, or a blue car may register as a false negative. It found a correlation that fits a pattern without fitting the pattern we wanted.
Unsupervised learning is the opposite of supervised learning in the sense that instead of assessing the accuracy of the training, you’re assessing the results of the process. You determine a set and supervise the process in getting the results with supervised learning, but you feed a set and let the system classify it to determine if a novel input is a member of a derived set or not.
This type of self-organization can lead to some interesting results. You let the system sort things out in a way that makes the clusters make sense based on the requirements. For instance, if someone handed you a bunch of vegetables and asked you to sort them, and you couldn’t tell what was what or the purpose of the sorting, you might go off of color. Potatoes, carrots, etc. all have different colors, but are the same thing, though your sorting method might disagree. You’re providing data and letting the system make sense of it based on some vague initial premise implied in the data.
This approach is useful in something like a medical application. If you pass a bunch of similar cell slides to an algorithm, it might find something which then correlates to cancer which can be explored further. While with supervised learning we gave the system our desired categories for output, here we give the system our inputs and let it make sense of it. You might find something that you didn’t even know to search for as an output with this sort of system.
Reinforcement learning is where the process of learning itself is fed back into. As your algorithm explores its environment, it learns more and tries to accomplish some specific goal(s) it is rewarded for. The better it does at approaching the goal, the more it is rewarded. Your algorithm reacts to the environment it is in to grow and adapt to fit the rules of the system.
A real life example of this is teaching a dog to walk on a leash. You don’t want the dog to pull or fall behind, but the dog really doesn’t understand the rules. When the dog does what it’s supposed to, you give them a reward and walk. When they don’t, they don’t get to continue walking or similar which shows them they’ve violated the rules (ideally you don’t unreasonably punish the dog since they don’t understand, but it’s fine to do to an algorithm). The dog learns the rules by trial and error and eventually knows instinctively what they can and cannot do when you have their leash.
Your algorithm is adapting to the changes as they come in rather than based on a static input set. This makes less sense for anything which doesn’t involve a continual process (a robot trying to walk and fighting gravity) or a dynamic environment (a video game algorithm where there aren’t necessarily fixed states). This learning system is based on trial and error rather than a fixed set of data.
Paradigms and Models
These 3 paradigms are the building blocks of other approaches. Supervised focuses on a task, unsupervised focuses on data, and reinforcement focuses on adaptation to an environment. None of these descriptions necessarily preclude the others. Supervised can be combined with unsupervised even though the names are opposite.
Machine learning has multiple paradigms, but it also has multiple models. While neural networks are the most common, they aren’t the only model. neural networks are arguably the most popular way to think of and approach machine learning, but the reason they work is the reason they aren’t the only model. You’re trying to reduce a complex biological process down to a mathematical one, so some tasks can be grossly simplified without a loss in accuracy for the model.
Artificial neural network were traditionally used to try and emulate the way a human or animal brain works. You have an input, you have an output, and you have some weighting process in the middle. This intermediate process is where data is combined or weighed to determine whether we have another dimension at play which affects the correct answer. An image is no longer just a group of pixels, but it’s a grouping of potential object boundaries each of which plays into the assessment.
Deep Machine Learning
This is part of the basis of deep machine learning. The added complexity in the middle makes the process a lot less black and white. The gray can mean more or less depending on the training and the basis of the model.
Deep learning requires extremely complex and robust training to be useful. If you train off of a limited set, you lose out on the effectiveness of the entire method. If you use pictures of cars on roads only and train off of it, the algorithm may end up looking for those roads as much as it looks for the elements of a car. There has to be a contrast to show what is in the set and what isn’t.
The more patterns you have in the data, the harder it is to not include them as elements of the task at hand. This is the primary complaint and general limitation of deep machine learning. We aren’t just having to deal with a much more complicated system, it’s also more fragile depending on the initial training. Limited training creates an artificially limited system.
The Paradigms in Action
By understanding these three basic paradigms, you can understand the more complicated systems which lead to more advanced learning paradigms. You don’t just have supervised or unsupervised, there can be a spectrum of adherence to one or the other. Supervised benefits task based metrics, while unsupervised can make sense of more input than classifications. Their names imply an opposition, but that opposition is on the basis of purpose rather than implementation.
There are countless other ways to approach each problem, but these three show you the main ways. The other paradigms are arguably derivative, niche, or else require understanding of the principles from the primary three. Orange is made from mixing red and yellow even though it exists as its own concept. White and black can also be thought of with respect to colors, and those colors are typically thought of in terms of the primary colors.
The paradigm is as much an axis as the model. What the goal of your task is affects how you should best teach your algorithm to learn. A task is different from raw data which is different from an environment. None of these preclude the others and can be combined to make more and more complex paradigms which suit more specific and specialized situations.