It comes from bigger niche called learning the representation of the data and can be thought as the collection of layered neurons. The input layers take a set of input, process it and pass the processed data to the next layer, which gives it to the next layer and so on until the terminal reaches the outermost layer. It is used to learn about unlabelled data. It has applications in automatic speech recognition and image processing.
The network layering looks as follows:
It is a type of learning algorithm wherein each action is rewarded with feedback if the action is good. For example, if a robot used for cleaning house, cleans the house then it receives positive feedback while if the robot messes up the kitchen then it gets penalized. It receives feedback from it action and the agents learn from it. It is different from deep learning in the sense that information isn’t moved forward but it rotates with an additional action. It is derived from our own psychological behavior.
It is thought to be a guidance to make better human decisions with help of software that imitates human behavior. IBM Watson is trying to make it so interactive that they can cater user according to their needs rather than giving them everything while intelligently communicating with the back end. It has application in speech recognition and sentiment analysis.
One cool example of cognitive learning is as follows:
Try to google ‘President of the united states’. Now, query ‘His wife’. It will understand the context and provide the answer accordingly.