Introduction to Neural Networks
U. of Minnesota, Final Study Guide
Psy 5038
Fall, 2003
There may be some material from the first half of the quarter, but the final
exam
will be heavily weighted towards the second half of the quarter.
Sample short answer questions
Define and describe the relation of the following key words or phrases to
neural networks. Provide examples where appropriate.(Answer 8 out of 12 items
drawn from set below; 3 points each).
"Energy" | attractor | pseudoinverse | bias/variance dilemma |
autoassociator | topographic representation | grandmother cell | asynchronous update |
Content addressable memory | Oja's rule | principal components analysis | sparse distributed representation |
constraint satisfaction | nearest-neighbor classifier | "explaining away" | correspondence problem |
gradient descent | Lyapanov function | encoder network | topology-preserving map (Kohonen) |
simulated annealing | cortical maps | generalized delta rule | Bayes net & probability factorization |
XOR | Hopfield's continuous response model | Gibbs G measure (Kullback-Leibler distance) |
anti-Hebbian |
spontaneous activity | projective field | receptive field | coarse coding |
marginalization & conditioning | radial basis function | prototype/exemplar | local minimum |
Sample essay questions
(choice of 2 essays drawn from a subset of those listed below; 12 points
each).
What is the Perceptron model? Discuss both its successes, failures and impact
on the field of neural network research.
Discuss the pros and cons of distributed vs. localized representations with examples from theoretical considerations and neurophysiology.
Give an account of Hopfield's 1984 graded response neural network model. How can it be used for memory? How does it relate to the discrete stochastic model of 1982?
Describe how Hopfield's 1982 neural network can be set up to solve a constraint satisfaction problem. Use an example, such as Marr and Poggio's 1976 formulation of the stereo problem.
How does the error back-propagation model work (you don't need to derive the learning rule)? What are the pros and cons of this learning algorithm?
Describe the Boltzmann machine algorithm for both recall using annealing, and for learning (you need not derive the learning rule). What are the pros and cons of this learning algorithm?
Give an account of just one of the following approaches to self-organization: Kohonen, 1982; or principal components sub-space extraction.
Discuss the problem of data representation in biological systems using as an example a known sensory or motor map (e.g. a tonotopic or topographic map).
What is a mixture model? How can EM be used to estimate the parameters of the model?