Module 2 of Machine Leaning Notes

MODULE-2 PARC Learning * Probably Approximately Correct Learning * In computational learning theory, it is a framework for mathematical analysis of machine learning algorithms * Make acceptable hypothesis with atmost & error. D = dataset for (i=1 to 100) ↑← Traindata sample (D) he Train (1) errorrate & error (h, Test) if errorrate > & k=k+1 if k<&8 100 PAC learning possible Failure (8) If error rate of all the test samples are greater than 8. Definition of PAC Learning: *het 'x' be an instance space and 'c'is a concept class for x, and ti 18 a hypothesis satisfying c and 'F' is a probability distribution function x; and the concept class 'c' us pac learnable, if there is an algorithm A which for sample drawn with am probability distribution F and any concept CEC with high probability produce a hypothesis in, whose ечном is small. Example for PAC Learning The rectangle 'abed' represents the hypothesis space such that it will bound all the positive instances tightly. So, we can say that the rectangular hypothesis make this algorithm classify mstances, tve or -ve with minimum error. 15/9/18 Noise * Noise is an unwanted anomaly in the data. * Noise may arise due to several factors: 1. Input attributes: Relevant attributes of a problem. 2. errors in labeling data - Misleading concept 3. Errors in instance selection *Noise leads to distortion of data, faulty predictions, so much training time, etc. Learning Multiple Classes Data set can be labelled to multiple concepts. Then such concept problems are called multiple classifications. Binary Classification > classification involving & labels. Here in our data set we have three labels A, B, C. Then to do a training we have to go for either binary classification multiple times or a design which can suit the classification in a better way. Binary classification multiple times Model Selection Model: * It is a mathematical function 01 graph or tree or any other structure which created using an algorithm that performs the specified conceptualization (labelling). *Model can be treated as a hypothesis * for a simple classification straight line is a model Inductive Bias * In a learning problem, we have only data. But the data may not be sufficient to find the solution. *80 on that occassion we can use some assumption (prejudice) to get the exact solution. The assumptions are called bias. Simple Models 1. easy to use 2. easy to train 3. Easy to explain 4. Better than complex model. Generalization * A model trained on the training set predicts the right output for new instances is called generalization. * Generalization refers to how well the concepts learned by the machine learning model apply to specific examples which are not been by the algorithm in past. * Two problems associated with generalization: 1. Underfitting: + is the production of a machine learning model that is not complex enough to acurately capture the relationship between features. 2. Overfitting It is the production of a machine learning model that is a complex function that try to capture the relationship between attributes features Testing Generalization Cross validation *hle can measure the generalization capability of a model, quality of the inductive base bias using testing. Here we will dake test case from training set (randomly) and new test cases (validation set). Dimensionality Reduction *An learning problems, data set have large number of features. * some times all the features are not that important. *20, we can apply dimensionality reduction techniques to reduce the number of dimensions.