Machine Learning is the ability of a computer to decide what a thing is by analyzing patterns in data and putting a name to them. For example, let’s say you wanted to teach a computer to distinguish a cat from a dog. You would keep feeding it labeled pictures of each species until it was able to tell the difference between the two without the label. This is called supervised learning, and it is one of the ways in which toddlers are taught. Hold up a flash card, point to the label, and say the word. Do it enough times and the child will eventually point to it themselves and say, “cat” or “dog”. In order to simulate this kind of cognitive functioning in machines, we have to create an artificial neural network (a special algorithm), train it to recognize patterns consistent with cats and dogs, then use those patterns to perfect the algorithm and categorize subsequent pictures accordingly. In AI speak, this is known as the supervised binary classification of linearly separable patterns, or a perceptron.
In addition to supervised learning, there is also unsupervised learning. This can be likened to a child that sees things in the world around them and is able to make connections with things they have already learned in the classroom. For example, let’s assume that a child knows what a cat is. Further, let’s assume the same child understands that houses and high rises belong to the same parent category, buildings. When the child sees a lion for the first time, it can generalize the concept of scale and apply it to cats . This is known as associative knowledge. Another common type of unsupervised learning is called clustering. A good example of this is predicting trends in home sales. In order to do this effectively, a computer needs an abundance of information including comps, square footage, acreage, number of rooms, architectural features, demographics, and so on. This data is then used to create patterns or sale clusters across the various attributes over time. As a practical matter, it too much data to compile under supervision. Therefore, a machine must mine the internet for these patterns in an unsupervised manner and decide if the results are relevant to the task at hand.
Finally there is reinforcement learning. Here, the aim is not so much the accumulation of knowledge, but rather the goal itself — to reach the highest score, to make the most money, to achieve the greatest efficiency. Goals are tied to actions, and trial and error is used to determine which action is best. The algorithms are self-teaching in that they are rewarded with points for exceeding targets and punished for not reaching them soon enough by taking points away. For example, high frequency trading uses artificial neural networks to make thousands of trades with modest gains over a short period of time. When the network is being trained, different trading strategies are tested to see which takes the least amount of time to meet or exceed the expected profit margin. The strategy or action with the most points is the one most likely to yield the best outcome. Further, once it goes live, the algorithm continues to train itself, adding more weight to the strategies that make the most money.
These are the three main methods of acquiring knowledge in AI, though you may have heard of another called deep learning. This is more of an ambiguous catchphrase than anything else. If you google it, you will find answers like a machine learning technique that teaches computers to learn by example; a subset of machine learning that uses vast volumes of data and complex algorithms to train a model; and a form of machine learning that is inspired by the structure of the brain. There is however one definition from IBM that makes some sense: “…a subset of machine learning, which is essentially a neural network with three or more layers.” Remember the perceptron mentioned earlier? For purposes of this discussion we can call the perceptron a single layer. This is fine for a simple application that just needs to tell the difference between cats and dogs. However most real-world applications require multiple layers of perceptrons to make decisions. Thus deep learning is a function of the complexity and number of layers in an algorithm, rather than a discrete learning methodology.
So which method is best? Each has its pros and cons. With supervised machine learning, you can control the data being sampled and therefore control the outcome, but the machine cannot teach itself. While this method excels at tasks like classification, the accuracy of the results is solely dependent on the samples provided which can be daunting to compile. Conversely, unsupervised machine learning reveals patterns in the data that humans might otherwise overlook due to volume or personal bias. Since these algorithms are self-teaching and there is no set sample size, learning is virtually limitless. While unsupervised methods excel at such tasks as predictive analysis, outcomes are difficult to reconcile due to the lack of training data. Reinforcement machine learning is also self-teaching in that actions are optimized through the process of trial and error. While this method excels at goal-oriented tasks such as stock trading, reinforcement algorithms carry more computational overhead, and too much reinforcement can be counterproductive (there is a point of diminishing return).
Of the three, unsupervised machine learning is both the most versatile and the most volatile. When implemented successfully, the results can be astonishingly effective, like with OpenAI’s ChatGPT. When implemented poorly, the results can be disastrous. Take for example Microsoft’s TAY, a conversational chatbot that was designed to simulate the personality of an American teenage girl. It was shut down within sixteen hours of its release for making highly offensive xenophobic remarks that were not only inconsistent with the intended persona but had no social filter whatsoever. More recently, Meta’s Galactica, a conversational chatbot purporting expertise in scientific knowledge was squashed after it received harsh criticism from the academic community for generating fake and misleading information. The lesson here is while big tech is rushing to market with advances in AI, caution must be taken so that machine learning does not distort the truth.
— — — — — — — — — — — — — — — — — — — — — — — — — — — — — — —
The following articles may also interest you: