CNNs
	Describe the parts of CNNs
	Describe the various ways that we can leverage a pre-trained CNN (for example, transfer learning).
	Describe specifically what is learned from the training set, validation set, and test set.
	Describe how CNNs relate to the "features + classifiers" approach to machine learning that I presented on the first day of class.
	Describe in words and pictures how comparing the accuracy on the training and validation sets during training can be used to detect overfitting and thus determine the optimal number of epochs to train the CNN to keep it from overfitting.   	

k means
	How does unsupervised learning differ from supervised learning?
	Describe the steps of the K-means algorithm.
	What bad thing can happen when the random initial cluster centers are chosen randomly?	
	Run the algorithm by hand on data using spatial distance (x,y)
	Predict the results on data using spatial distance
	Predict the results on data using color distance (r,g,b)
	Predict the results on data using both color and spatial distance together (5 dimensions: r,g,b,x,y)
	(For example: How does the introduction of spatial (XY) distance improve clustering done on RGB only? If XY are weighted too heavily, what is the danger?)
	
Template matching
	Discuss the idea of template matching
	What are the issues with basic template matching? What can we do to overcome that limitation
	How is template matching implemented using a filter?
	Describe how cross-correlation can be used to match templates.
	Describe HoG features.

Hough transform
	Describe the principles used by the Hough transform
	What does populated voting space look like in ideal and non-ideal cases for line finder? Circle finder?
	What parameters would you use to search for a line? A Line segment? A rectangle? Other shapes?
	Draw the parameter space showing the votes that are cast when searching for a line, segment, square, rectangle, etc.
	

	
PCA
	How do you compute the principal components? (eigenvectors of what matrix?)
		What dimensions are the various matrices? 
	How do you project data onto those principal components? 
	How does one reconstruct a data point back into the original coordinate space?
	What are the principal components (visually)?
	How do you determine what dimension you can reduce your data set to?
	When one performs PCA on a series of nature scenes, the first principal component corresponds to which color axis? The second and third?
	Describe how the relative sizes of the eigenvalues relate to the distribution of the data in the original space.
	Novel applications of PCA, showing you understand the process.
	What are eigenfaces?
	Discuss how PCA was applied to derive the LST color space.

	
	
Cumulative:
	Can you explain your term project or past projects in terms of the features + classifier framework I presented on the first day of class?

	
Other (chosen from term to term - we usually run out of time for most of these topics):


Bayesian classifiers
	Given distributions of features for two classes:
		Find the optimal threshold (the one that minimizes the error rate) and justify why it must be there.
		Draw ROC curves that arise from moving the threshold.
	Describe classification using the MAP principle.
	
	Use the MAP principle to find the most likely class, given priors and evidence from one or more detectors.
	Interpret an intensity histogram and compute an optimal threshold from probability density functions of the foreground and background.


Motion:
	Compute motion vectors.
	Draw vector fields for pan/zoom/etc.
	Discuss the aperture problem. Show where it arises. Describe a way to solve it.
	Show how one can use a Taylor series to determine a motion vector (or set of motion vectors) that is consistent with a given point. Explain how we know each of the variables in the expansion (f(x,y,t), df/dx, di/dx, etc)
	How does one find interest points?
	Why is tracking interest points faster than finding them initially?

Describe the concept of boosting (Adaboost).
Kalman filtering (for motion)
Normalized cuts (for image segmentation)