CMPS 242 Midterm Topics (Fall 2005)
(Version 1)
The midterm will be in class Tuesday Feb. 22. The exam will be closed-book, although
students may have one 3x5 card of handwritten notes (both sides). Here are the topics we have
covered so far in CMPS 242 this quarter:
- What
is Machine Learning
- Classification
and Concept learning, hypothesis class and version spaces
- VC
(Vapnik-Chervonenkis) dimension and shattering points by hypothesis class
- Perceptron
algorithm and Widrow-Hoff gradient descent techniques for learning
halfspaces
- Logistic
regression
- Linear
discriminant analysis
- Noise
and causes (label errors, attribute errors, features or hypothesis class
may not fit phenomina exactly)
- Overfitting
– Using test sets to detect reduction in generalization
- Basic
probability (sample space, events, random variables, independence,
conditional probability, Bayes rule)
- Estimating
probabilities (e.g. the bias of a coin flip): maximum likelihood, priors,
maximum aÕposteriori, mean aÕposteriori, and Laplacian estimates of
probability
- Hypotheses
as models generating the data, identifying the maximum likelihood
hypothesis, the maximum aÕposteriori hypothesis, and the mean aÕposteriori
predictions
- Least
squares regression as Maximum likelihood (assuming Gaussian noise)
- Asymmetric
losses, Bayes risk, and Bayes optimal predictions
- Decision
boundaries and discriminant functions
- Learning
Gaussian distributions – biased and unbiased estimators of the
variance
- Gaussians
with same co-variance matrices give linear decision boundaries
- Bayes
Nets (graphical models, belief networks)
- Na•ve
Bayes algorithm
- Decision
Trees: Greedy construction of
trees, information gain criterion, applying a split criterion (impurity
function, e.g. information gain) to select tests at nodes, overfitting and
pruning
- Feed
forward artificial neural networks, training using backprop algorithm
(gradient descent)
- Nearest
Neighbor algorithm and variants.
Instance based density estimation.
- Support
vector machines and Kernel functions
- Boosting
Recommended Problems: