About ML

Intro

Machine Learning is the science (and art) of programming computers so they can learn from data.

Or in a more engineered-way:

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.

By Tom Mitchell, 1997

Types of ML

Machine Learning works on maths. Ranging from the very basic to the very complex.

There are various types of Machine Learning algorithms. Let’s start with the Supervised Learning.

Supervised Learning

In this type of learning, you provide questions along with their answers to the machine. Then, the machine first looks at the question and then at the answer and then tries to understand the relationship between them. The question and the answer forms the training data. Sufficient amount of training data enables the machine or algorithm to understand the relationship between the data and its solution.

Now, if you just feed a new question to the machine (related to the data initially the machine was trained upon; don’t expect it to run if you didn’t tell it how to walk 🙂), it will try to predict its solution based on its understanding of the training data.

Supervised learning involves many alogs:

• Regression
• Classification
• k-Nearest Neighbors
• Support Vector Machines (SVMs)
• Decision Tree and Random Forests 🌲🌳
• and the most popular: Neural Networks 🚀🚀 (they may or may not be supervised)

Example:

The most famous example: Predicting house prices given some info about the house falls under supervised learning (Regression).

Unsupervised Learning

Now suppose that you don’t provide the machine with the solutions. Instead you just provide it with the data, and say to it, “Learn it yourself!“. And that’s it. It’s a learn-it-yourself approach. The system tries to learn without a teacher or a solution. It tries to find the patterns and works out on its own.

It involves:

• Clustering

  • k-Means
  • Hierarchical Clustering Analysis (HCA)
  • Expectation Maximization

• Visualiztion and dimensionality reduction

  • Prinicipal Component Analysis (PCA)
  • Kernel PCA
  • t-distributed Stochastic Neighbor Embedding (t-SNE)

Example:

Say you have a lot of supermarket data. You can run Unsupervised algorithm to discover interesting relationships. Like you may find that people who buy Item A tend to buy Item B too. So, you may want to place Item B close to Item A.

Semi-supervised Learning

This is the mid-way between the previous algorithms. Just train your machine on a lot of unlabelled data and a little bit of labelled data.

Example:

Take Google Photos for example. It automatically recognizes same faces in different photos without you telling it. After it recognizes that these persons appear in these many photos, you just need to tell it who those persons are, just one label per person and it will automatically be able to name everyone in every photo. Useful for searching, eh?

Reinforcement Learning

It’s a totally different thing. Think of an agent (can be anything) that observes the environment, selects and performs actions and get rewards in return or penalties in the form of negative rewards. It must learn by itself what is the best way (called policy) to get most rewards over time.

Example:

Suppose you have a robot. And you have water on one side and fire on other side. Now, the robot moves and observes it’s environment. Then, it performs an action using a strategy. Say, if it moved towards fire (and fire is said to cause problem 🙂), it may get damaged. It gets a negative reward for getting near fire. So, next time it will try to not go near fire so that it can maximize it’s reward.

---

So, this is the basic overview of ML. You give data to the machine, train it on the data and then make predictions or other insightful decisions.