Machine Learning 101
Machine learning formal definition.
Arthur Samuel (1959): Machine learning is a “Field of study that gives computers the ability to learn without being explicitly programmed.”
My intuition: Arthur thought that when we program, we write logical syntax which allows the computer to do the task, however, machine learning is a field where this logical syntax would be learned by the computer itself through observation of patterns of dos and don’ts and minimizing the loss function.
Tom Mitchell (1998): According to him “A computer program is said to learn from experience E with respect to some task T and some performance measure P, if its performance on T, as measured by P, improves with experience E.”
My intuition: Basically, the computer gets an intuition of how to perform the task by repeatedly doing it and getting some rewards for it, the more it performs the task successfully it becomes a good learning agent.
Types of machine learning
Get romesh prasad’s stories in your inbox
Join Medium for free to get updates from this writer.
Supervised learning
Informal definition: Given a labeled cleaned dataset, the goal in this is to map labeled input to labeled output. If we look at Tom’s definition, the task (T) would be understanding the relationship between the output and input, and performance measure P will be how good this mapping is. So once it is learned through training, the learning algorithm, given a new input should be able to map it to a close approximate output.
The simplest way of achieving this mapping is by passing a straight line (linear regression) or curve line (quadratic regression). This type of technique is called regression. If the output data has binary values (1,0) or multiple classes (1,2,3,4), then the mapping between input and output is achieved through classification methods such as logistic regression. Here the input labeled data are often called Xi, where I can be 1,2,3,..,n, and output data are called Y. Mapping between Xi and Y ( Xi~Y).
Unsupervised Learning
Informal definition: Given input labeled data but no output data the learning algorithm is asked to find patterns that are not easily visible through the naked eye. The pattern found is often done through clustering. Examples: social networks, finding similar content from different creators and clustering them together, differentiating between different audios if given input is noisy audio, etc., Basically, the learning algorithm is asked to find something interesting given a labeled input only.
Reinforcement Learning
Informal definition: This is very different than the other two learnings. In this, we don’t have the inputs and outputs. Here, we have an agent whose objective is to take action in a state of the environment and receive a reward for each action. Based on the face value of this reward the agent perceives good behavior versus bad behavior. The goal of the agent is to maximize good behavior. So, over time the agent should be maximizing cumulative reward. Example: Not every system or joint movement of robots can be mathematically modeled to do a task, additionally the environment for the robot is our nature, which is not a deterministic environment. Hence an implementation of an RL agent controller to control the robot can allow the robot to be generalized over different problems and also find optimal control in some of the problems.
