Question

Artificial Intelligence and Machine Learning
- How is traditional programming different from current programming in the sense of ML (machine learning)?
- Explain the two components of a learning problem. What is learning a hypothesis?
- When to use a simple hypothesis compared to a complex one? (Underfitting vs. Overfitting)

- What is the difference in the error in training and testing and how does this relate to the generalization of the learning algorithm?

Answer 1

a.In traditional programming,the programs needs input and the logic which should be hard-coded for it to give the output.Whereas in machine learning,it all needs some data from which it develops a logic and using this logic it gives output for the new data which means in machine learning most of the part is done by the machine itself i.e interpreting the logic from the given data and applying that on the new data but in traditional programming the user himself should provide both input and the logic for it to produce output.

b.In machine learning the two components of learning problem are as follows:
1.Supervised Learning
2.Unsupervised Learning
1.Supervised Learning : In this the learning is supervised.The data(input and output) is given with labels so that it can use either of "Classification" or "Regression" techniques to understand and develop the logic.Firstly,it will be trained with some set of data and then it needs to work on the new data.
2.Unsupervised Learning :In this learning is not supervised.The data(input and output) is given without any labels and everything is on the machine itself to develop the logic.Here the machine wont be trained properly and when we give some new data the output produced may not be satisfactory in all the cases.
The learning hypothesis is nothing but the evaluation of accuracy in the program.It may not pass in all the cases but it should be in such a way that one can test it.This consists of the mapping between the input and output.

c.Simple hypothesis can be preferred when there is one dependent and one independent variable exists and also the relationship between both of them exists.In complex hypothesis there exists too many variables i.e more than one dependent and independent variables.So, when the algorithm or program is simple it uses the simple hypothesis over complex hypothesis.In both underfitting and overfitting the results wont be up to the mark as expected.

d.If a learning algorithm performs well with the training data and performs low with the test data then we cannot generalize the learning algorithm.The difference should be very minimal between the training and test data for a learning algorithm to be generalized i.e it should perform well with both training and test data.This scenario is known as "Overfitting" where in the training data is of very high standards when the new data doesn't require that standard.So,for the generalization of a learning algorithm the difference in the error between training and test data should be very minimal or negligible.