The ability to learn is essential for any system that claims to possess intelligence. Field of Artificial intelligence is no different. The machine learning technology in AI enables computers with the ability to learn without explicitly programmed to do so. Recommendation engine used in the e-commerce portal, fraud detection in financial sector, sentiment analysis in social media are few examples that use machine learning in our day-today activities.
Over the last 50 decades, scholars have come up with many algorithms that enable the machines to learn. The learning style used in these algorithms are broadly categorized in four types - “Supervised learning”, “Unsupervised learning”, Semi supervised learning” and “Reinforcement learning”. Let’s try to understand these better.
Supervised learning:
This style of learning requires human interference. Its like teaching the machine “if you are using 1 cups of milk, 1 cups of water, I tea spoon tea leaves and 2 tea spoon of sugar, you get two cups of tea.”
Subject matter experts provide the manually tagged historical data (big chunk of data), referred to as “training data”. This training data typically consists of “input” (ingredients in the above example) and “desired output” (Two cups of tea in the above example). This is called “Labeled data”. Machine is then required to come up with a relation between the inputs that result in the desired output. This relation is referred to as “model”. Based on this model, machine is then asked to predict the output for the given new set of input data.
Application area
- Classification Problems
- Regression problems
Example
- These days supervised learning is widely use in cancer prognosis. Here huge number of medical records relevant to cancer is collected and used to train the system. When a new medical case is given, then system tries to predict if a tumor is malignant or benign based on its size and other defined factors.
- Nounshoun application has trained its system to predict parts- of- speech in any given sentence. The model is trained using a big set of hand tagged sentences.
Algorithms
- Here are few well known algorithm that use this approach:
- Logistic Regression
- Convolutional Neural Network
- Support Vector Machine
Unsupervised learning
Here the training data fed to the system is not labeled. Machine is required to analyze the data and group them by their similar features. These are mainly used to discover structure in unlabeled data.
Application area
- Clustering problems
- Dimensionality reduction problems
- Association rule learning
Practical Example
Recommendation systems on e-commerce portals and social media use unsupervised learning methods. Bases on users browsing history, they first group people with similar interests, and then give same suggestions to people in same group.
Algorithm
- Here are few well known algorithm that use this approach:
- K-means clustering
- Density based EM algorithm
- Hierarchical clustering
Semi-supervised learning
This approach it lies in between supervised and unsupervised learning. Here the training data is combination of small portion of labeled data and major portion of labeled data. Here systems are expected to predict the output as well as group/organize them to do better predictions.
Obtaining huge amount of labeled data can be very expensive. Semi-supervised learning approach proves to be beneficial both in terms of cost reduction and better accuracy.
Application areas
- Classification Problems
- Regression problems
Example
- A good example of semi-supervised learning is face recognition technology that Facebook uses. Using its enormous picture data, it has trained the system to recognize the facial patterns in a picture and figure out faces in it.
- Computer vision is another area where semi supervised learning is widely used.
Algorithms
Here are few well known algorithm that use this approach:
- Generative models
- Bayesian Network Classifiers
- Graph-based methods
Reinforcement learning
In a given environment, machine is made to learn the ideal behavior by giving reward/positive feedback for the correct actions and punishment/negative feedback or the incorrect actions. This method is called reinforcement. Here, the goal for the system is to gain the maximum rewards and thus learn to behave in the desired manner.
Application area
- Robotics
Examples
- Google’s famous self-driving cars use reinforced learning to improve knowledge of routes.
- An awesome use case for reinforced learning I came across is Lilly drone camera. You’ll be amazed to see how this drones self- learns to fly and then with the help of sensor follows you to capture you video.
Algorithms
- Here are few well known algorithm that use this approach:
- Q-learning
- Sarsa
- Actor-Critic method