Learning Curves

  |   Source
Learning Curves
  • Curve to check the learning algorithm, whether bias, variance, or both

  • To plot learning curve, take small subset of training set(10-40 training examples)
  • training set for m = 3, training set = 0. slightly larger if using regularization
  • Learning curves will only make a shape only based on that three training example
  • Say m >= 4, will become harder for the curve to shape all training set perfectly
  • Concretely more training examples, the harder the curve to fit all the examples.
  • The more data we have, the more fit the cross validation error, that is better hyphothesis for more training examples
  • Sooo, with m start small, the error of Jtrain will be small, because it will overfit smaller data
  • This is what we get for the curve if what we have is either bias or variance
  • First choosing the model (See hypothesis formula) that is only the linear model.
  • The high bias can be shown as the three Jtest, Jtrain, and Jcv are all end up at high value
  • suffering High Bias can't really help much if we get more training data.
  • As we taught earlier, the Jcv will be similar to Jtrain if the learning suffer high bias, the error will be usually high.

  • Next is if we  suffering high variance, denotes by Jc) >>(much higher) Jtrain.
  • Suppose we use hypothesis with a lot of high order polynomials and small lambda
  • Jtrain if get more data, will generally close to future example,
  • But more the data the cross validation will keep decreasing(less error), and the Jtrain will keep increasing(more error)
  • Sometimes curve can be a lot cleaner/messier, but still helps out intuition if it has bias/variance problem

  • Next, discussed about spesific actions to take (or not take) to increase the performance of learning algorithm