Inductive learning and deductive learning are two approaches to machine learning. Inductive learning involves learning from examples, while deductive learning involves learning from rules and knowledge.
Hi … I have a question why the need of a lot of data in larger hypothesis space maybe I could use the same exact data for all the hypotheses in the hypothesis space ?
Because certain hypotheses within the broader hypothesis space may be overly complex. For instance, within a hypothesis space consisting of polynomials, increasing the degree also increases the complexity and expressiveness of some hypotheses.
Hi! Has dimensionality reduction been discussed in any of these lectures, PCA and t-Sne in specific? If yes, please share the lecture number. I didn't see any lecture name targeting this topic.
Great Lecture! I have a question: in K nearest neighbor, is it possible that the test accuracy will be larger than training accuracy as K increase? Thanks!
It could happen, but that indicates problem. Either it is a fluke on the test set(could be caused by not enough data in the test set) or the hypothesis h trained is underfitting. Correct me if I'm wrong, this is my understanding.
Absolutely! Basically if K tends to "infinite" (when K = the number of elements in the training set or test set), then the prediction would be just the average (for regression) or the mode (for classification which I guess is the one your are asking) of all input set. If the test set just happens to have more samples that belong to that mode class then the accuracy of testing will be greater than during training.
@@albertmiao1976 Yes, if this happens usually it means that you are underfitting which is the case when K is to big in K-NN (K-NN tends to a constant classifier if K is to big)
Patient lecturer! In K-fold validation, average accuracy is reported. So, there is no testing set? Or, the average accuracy is reported as training accuracy? In K-fold validation, you get K models. So, which model to use among the K models?
Thats the validation accuracy from where we choose the value of K. There is a separate testing set where we do not do training at all. We use to K value from cross validation and report the testing set accuracy as the accuracy of the model.
Day 1 complete. Amazing lecture. Thank you professor.
Understood well!
Thanks! Very straightforward
Amazing lecture! Thank you!
Inductive learning and deductive learning are two approaches to machine learning. Inductive learning involves learning from examples, while deductive learning involves learning from rules and knowledge.
This is awesome, thank you
Hi … I have a question why the need of a lot of data in larger hypothesis space maybe I could use the same exact data for all the hypotheses in the hypothesis space ?
Because certain hypotheses within the broader hypothesis space may be overly complex. For instance, within a hypothesis space consisting of polynomials, increasing the degree also increases the complexity and expressiveness of some hypotheses.
Hi! Has dimensionality reduction been discussed in any of these lectures, PCA and t-Sne in specific? If yes, please share the lecture number. I didn't see any lecture name targeting this topic.
checkout CS 485
Check out MIT lectures for that.
Great Lecture! I have a question: in K nearest neighbor, is it possible that the test accuracy will be larger than training accuracy as K increase? Thanks!
It could happen, but that indicates problem. Either it is a fluke on the test set(could be caused by not enough data in the test set) or the hypothesis h trained is underfitting. Correct me if I'm wrong, this is my understanding.
Absolutely! Basically if K tends to "infinite" (when K = the number of elements in the training set or test set), then the prediction would be just the average (for regression) or the mode (for classification which I guess is the one your are asking) of all input set. If the test set just happens to have more samples that belong to that mode class then the accuracy of testing will be greater than during training.
@@albertmiao1976 Yes, if this happens usually it means that you are underfitting which is the case when K is to big in K-NN (K-NN tends to a constant classifier if K is to big)
Patient lecturer!
In K-fold validation, average accuracy is reported. So, there is no testing set? Or, the average accuracy is reported as training accuracy?
In K-fold validation, you get K models. So, which model to use among the K models?
Thats the validation accuracy from where we choose the value of K. There is a separate testing set where we do not do training at all. We use to K value from cross validation and report the testing set accuracy as the accuracy of the model.