Summary of topics in the course

K-nearest neighbour, average vote for K nearest training samples.

Decisions trees: Entropy = unpredictability = sum -p log p. Maximize information (= - entropy) gain for every node. (gini = - sum p(1-p). Crossvalidate and prune the last nodes.

Bayesian Inference: The process of calculating the posterior probability distribution P(y | x) for certain data x.

Bayesian Learning: The process of learning the likelihood distribution P(x | y) and prior probability distribution P(y) from a set of training points.

Liklihood: P(x|y)

a posterori: P(y|x) proportional to P(x|y) P(y)

Boosting: Aggregating many weak classifiers. Adaboost: Weight misclassfiied samples iterate, and then use weighted average of classfiiers.

Bagging: Bootstrap new samples, trains classfiers and averge them.

Concept: c true/false labeling of x in X

Hypothesis space: All possible true/false concepts, h in H (before data arrives)

True error of hypothesis: probability that hypothesis h gives wrong classification for one datapoint.

Probably Approximately Correct: How many training samples are needed if we want probability that any hypothesis missclassifies possible data to be less than delta = H*(1-eps)^m

VC dimension: The largest set of data for which each subset can be described by a hypothesis h in H.

Naive Bayes: Assume features are not dependant, maximise aposterori (not liklihood)

Logistic regression: Regression to a probability.

Inference and decision:

Discriminative function

Discriminative vs Generative model

Credit Assignment - The problem of enforcing the right parts of a compound behaviour