What Exactly Machine Learning?
What Exactly Is Machine Learning?
Here’s the simplest definition I came across, from Berkeley: Machine learning is “[…] the branch of AI that explores ways to get computers to improve their performance based on experience.”
Let’s break that down to set some foundations on which to build our machine learning knowledge.
Branch of AI
Artificial intelligence is the study and development by which a computer and its systems are given the ability to successfully accomplish tasks that would typically require a human’s intelligent behavior. Machine learning is a part of that process. It’s the technology and process by which we train the computer to accomplish a given task.
Machine learning techniques are still emerging. Some models for training a computer are already recognized and used (as we will see below), but it is expected that more will be developed with time. The idea to remember here is that different models can be used when training a computer. Different business problems require different models.
Get Computers to Improve Their Performance
For a computer to accomplish a task with AI, it needs practice and adaptation. A machine learning model needs to be trained using data and in most cases, a little human help.
Based on Experience
Providing an AI with experience is another way of saying “to provide it with data.” As more data is fed into the system, the more accurately the computer can respond to it and to future data that it will encounter. More accuracy in understanding the data means a better chance to successfully accomplish its given task or to increase its degree of confidence when providing predictive insight.
- Entry data is chosen and prepared along with input conditions (e.g. credit card transactions).
- The machine learning algorithm is built and trained to accomplish a specific task (e.g. detect fraudulent transactions).
- The training data is augmented with the desired output information (e.g. these transactions appear fraudulent, these do not).
How Does Machine Learning Work?
Machine learning is often referred to as magical or a black box:
Insert data > magic black box > mission accomplished.
Let’s look at the training process itself to better understand how machine learning can create value with data.
Machine learning is dependent on data. The first step is to make sure you have the right data as dictated by the problem you are trying to solve. Consider your ability to collect it, its source, the required format, and so on.
Data can be generated by different sources, contained in different file formats, and expressed in different languages. It might be required to add or remove information from your data set, as some instances might be missing information while others might contain undesired or irrelevant entries. Its preparation will impact its usability and the reliability of the outcome.
Depending on the size of your data set, only a portion of it may be required. This is usually referred to as sampling. From the chosen sample, your data should be split into two groups: one to train the algorithm and the other to evaluate it.
As commonly seen with neural networks, this stage aims to find the mathematical function that will accurately accomplish the chosen goal. Using a portion of your data set, the algorithm will attempt to process the data, measure its own performance, and auto-adjust its parameters (also called backpropagation) until it can consistently produce the desired outcome with sufficient reliability.
Once the algorithm performs well with the training data, its performance is measured again with data that it has not yet seen. Additional adjustments are made when needed. This process allows you to prevent overfitting, which happens when the learning algorithm performs well but only with your training data.
The model is optimized before integration within the destined application to ensure it is as lightweight and fast as possible.
Are There Different Types of Machine Learning?
There are many different models that can be used in machine learning, but they are typically grouped into three different types of learning: supervised, unsupervised, and reinforcement. Depending on the task, some models are more appropriate than others.
With this type of learning, the correct outcome for each data point is explicitly labeled when training the model. This means the learning algorithm is already given the answer when reading the data. Rather than finding the answer, it aims to find the relationship so that when unassigned data points are introduced, it can correctly classify or predict them.
In a classification context, the learning algorithm could be, for example, fed with historic credit card transactions, each labeled as safe or suspicious. It would learn the relationship between these two classifications and could then label new transactions appropriately, according to the classification parameters (e.g. purchase location, time between transactions, etc.).
In a context where data points are continuous in relation to one another, like a stock’s price through time, a regression learning algorithm can be used to predict the following data point.
In this case, the learning algorithm is not given the answer during training. Its objective is to find meaningful relationships between the data points. Its value lies in discovering patterns and correlations. For example, clustering is a common use of unsupervised learning in recommender systems (e.g. people who liked this bottle of wine also enjoyed this one).
This type of learning is a blend between supervised and unsupervised learning. It is usually used to solve more complex problems and requires interaction with an environment. Data is provided by the environment and allows the agent to respond and learn. In practice, this ranges from controlling robotic arms to find the most efficient motor combination to robot navigation where collision avoidance behavior can be learned by negative feedback from bumping into obstacles. Logic games are also well-suited to reinforcement learning, as they are traditionally defined as a sequence of decisions, such as poker, backgammon, and more recently Go with the success of AlphaGo from Google. Other applications of reinforcement learning are common in logistics, scheduling, and tactical planning of tasks.