In any AI system, the hardest things to learn are obvious tasks, such as walking. Walking is an overlooked task because it's so natural to us--in fact, we struggle to even articulate what makes us walk the way we do.
Chess, on the other hand, is not an overlooked task. It has a specific set of rules that might not be easy for an average human being to understand or adhere to, but there is a certain logic to them. This fact makes it easy to teach a computer how to perform.
Consequently, the first step in training any computer to make intelligent decisions is to find the “chess” elements of a problem.
Another way to understand this is to classify all tasks into three categories:
The formal tasks are the ones that follow basic physics rules. We need a simple set of instructions to perform them. For example, how to pick a glass from a table. Formal tasks don't require any specialized skills or complex rules. Almost any human can do them with ease.
Expert tasks, on the other hand, are specialized problems with a complex set of rules. Not every human being is equipped or required to perform these tasks. Chess is a prime example, along with other games like Go. Because at a fundamental level they are a set of mathematical rules, you can teach computers how to perform these tasks on their own rather easily.
The most difficult ones are the mundane tasks, mainly because they are overlooked and obvious. Most humans are able to perform these subconsciously, which makes it harder to teach them to a computer. Walking is one such task, and most of us do it without much thinking. Consequently, it’s hard to distill the ground rules..and that's precisely the reason it's so hard to teach computers how to walk.
The Toughest Challenge
The hardest part for any learning algorithm is to identify the “chess” elements in each problem, or specifically, what the “math” is that’s underlying a problem. Then, we must try to teach that math to computers.
For example, in self-driving cars, one “chess” element is the geolocation data. Not to say that it's easy, but it's ultimately a mathematical/technical problem...and technical problems can be coded. The difficult part is to make a self-driving car differentiate between a toddler and a bus both crossing the road. However, machine learning is particularly good at solving the “chess” pieces of the puzzle.
Three Types of Machine Learning
There are three major types of machine learning that are used to solve these “chess” problems:
1. Supervised Learning
2. Unsupervised Learning
3. Reinforcement Learning
All three of these require data for training, and what sort of data that is and how we can use it to train a computer varies.
In supervised learning, we know the truth or the output, and we train a computer by telling it that truth via samples.
A typical example is email spam filtering. We know what a spam email looks like. Also, supervised learning is used where there is a definitive yes or no answer. An email is either a spam or it's not. A patient either has diabetes or doesn’t.
In supervised learning, we train a computer by showing what a spam email looks like. Eventually, it starts to figure out the symptoms itself. Supervised learning is the oldest type of machine learning, and its use is widespread in performing tasks that require a narrow level of intelligence.
In unsupervised learning, we don't know the truth or the output. Instead, we have to teach a computer to create a structure from the data itself. Identifying market segments from customer data is an example of unsupervised learning--we can't be sure of how many market segments are there.
Categorizing a large set of articles is another example. The categories can be any number and we don't know if an article is part of one, two or more categories. Unsupervised learning is tricky, and we are yet to achieve a major milestone using it.
Reinforcement learning sits somewhere in between supervised and unsupervised learning. You know the parts of the truth or output, but not the whole truth. Based on that you teach a computer algorithm to perform some action. If right, the action is rewarded; if wrong, the action is punished. Based on this reward system, the computer learns to know whether what it did was right or wrong. Consequently, learning propagates back into the system and becomes part of the knowledge system. The next action by the computer will be more informed. Eventually, we will get to the truth by taking actions, learning the effectiveness of the action and making adjustments.
In a way, this is how we learn about life as well. While we don't know everything, we make decisions based on what we know. The goal is not always to make the right decision, but rather to extend the limits of what we know about ourselves.
Weighing the Options
The choice of which type of learning machine learning algorithm to use is dependent on the task at hand. In general, supervised learning is great for forward-feeding classification problems like detecting spam emails. Unsupervised learning is yet to have a major breakthrough. Right now, reinforcement learning is where all the hype is. The nesting of multiple layers of neurons, hence the term “deep learning,” makes the back propagation of knowledge a true game changer. And is well suited to solve complex problems like self-driving cars.
Whatever type of machine learning is employed, it’s undoubtedly changing the way we approach almost everything. A variety of organizations are already benefiting from the automation of lengthy and repetitive tasks, as well as predicting outcomes for new data. Artificial intelligence, machine learning and deep learning are redefining the way we work, and by getting informed about the range of their applications, organizations (and individuals) can ensure they remain competitive and relevant in a rapidly changing society.
Interested in exploring the latest in AI and machine learning?
Don’t miss Machinery.ai, a one-day conference coming to Seattle and DC this November! Join other developers in exploring the digital machinery of a new era. Find out more information and RSVP at machinery.ai.