Learning Inference Simplified

Transfer Learning 

The attempt to use the output of one trained model as an input to another model which is to be trained is called the transfer learning.

Say We have trained one model to find animals. Then we have another model to find the endangered species.So with one set of images we train the first model to find the animals and the second model when we give the input image set we give these features along , which will help us to find the endangered species more accurately.

Inference 

Suppose we are getting continous samples from a source , say the price of a stock during an election. So the price of the stock would be continously varying with respesct to the prospect of one poltical part or other winning. Say we have this data. And we are able to fit a probability distribution over it. Means we are able to put a graph which data samples will follow in a similar case in future also. So we can use this to algorithmic trading during the next election and see some money (Hopefully). This process of figuring out a distribution statistically from data is called the Inference.

Reinforcement Learning

Most of Us would have heard about the Pavlov experiment with dog. When a dog is trained with reward to believe that a certain stimulus like a bell is associated with food the dog will salivate even without food when it hears the bell next time.Reinforcement is a method of strengthening a pattern of behaviour with a certain stimulus.In the case of Machine Learning, the behaviour we expect from the neural network model is giving us the targeted feature detection . For achieving this we have to design the network in such a way that the network has to strengthen the path way from input to output with respect to the right kind of input it is receiving. So a re inforcement learning network will have a a policy by which it will strengthen the network based on some input feature.The strengthening of the netwok (More precisely adjusting the weights ) is the reward which we are targeting with this approach. The input feature is our stimulus.Its more like a a viterbi decoder and neural network working together to achieve the results.