|
Week/Date |
Topic/Reading |
Deliverable(s)/Notes |
|
August
25 |
Introduction,
historical perspective, hard problems, sensory processing, the Turing Test |
Begin
readings from text—Chapters 1-3 |
|
August
30, September 1 |
Machine Learning Unsupervised Learning Markov chains ·
Building
a probabilistic state machine (PSM)
·
Using
a PSM to generate words
|
Chapter
6 Program Project 1: Create a word generator
for a language whose example words are {spare, spear, pares, peers, reaps,
peaks, speaker, keeper, pester, paste, tapas, pasta, past, straps, tears,
terse, steer, street, stare, rates, streak, taste, tapa,
peat, eat, ate, tea, seat}. ·
Use
your program to generate at least 100 random words. ·
Are any words repeated? ·
What is the probability of generation for each word? ·
What are the shortest and longest words generated? ·
What is the longest word that can be generated and what
is the probability of its occurrence? If you cannot answer this directly, can
you provide a bound for the probability of occurrence? ·
Use
the words generated by your program to create a second Markov model and
compare the transition probabilities of the two models you created. ·
Use
you program to create two Markov models based on 500-word (or larger) writing
samples from two, different languages (say, English and French, American and
German, or possibly American and English). The writing samples are to be
taken from web sites using different modern languages that are written using
Latin letters. Compare the transition probabilities of the two models you
create. |
|
September
6, 8 |
Machine Learning (continued) Nearest
neighbor classification |
Program Project 2: Extend your prior work by creating Markov models for three writing samples that employ different languages, not previously used (in your work for Program 1). Among the five models perform leave-one-out tests to identify the nearest neighbor among the remaining four for each language model. · What similarities do you expect? · What are the statistical properties of the nearest neighbor measurements for each test? · Were the nearest neighbor relations symmetric? ·
Construct and report a matrix that indicates
the similarity measure for each pair of models. |
|
September
13, 15 |
Machine Learning (continued) |
Due 13 Sep: Report of findings and in-class, demonstrations of Programming Project 1 |
|
September
20, 22 |
K- means
clustering (continued) |
Due 20 Sep: Report of findings of Programming Project 2 Program Project 3: Consider the two-dimensional data given in test file xyz.dat to partition it into three, four, and five clusters, using k-means clustering. Plot the data and the cluster centers for each partition. Repeat the process using different initial cluster centers to construct partitions into three, four, and five groups. Compare the clusters with the first partition |
|
September
27, 29 |
Bayesian Classification ·
Review
of basic descriptive statistics ·
Basic
operations with matrices ·
Extending
statistics to multiple dimensions ·
Covariance ·
Discriminant
functions |
Due 27 Sep: Report of findings and in-class, demonstrations of Programming Project 3 Program Project 4: Use the data given for PP3 to find the matrix form of the discriminant functions, assuming that the data are partitioned into three classes. Use the discriminant functions to classify data points provided in the file rst.dat into the various groups. Plot the original data, distinguishing among the three classes, and the new data, also distinguishing among classes. · Were the points classified as you expect? ·
Find and plot the discriminant boundaries
among the three classes? |
|
October
4, 6 |
Bayesian classification (continued), review for
Test 1 Test 1: 4 October 2011 |
|
|
October
13 |
Strategies to Find
Solutions for Search, Constraint Satisfaction, and Optimization Problems Evolutionary computation |
Due 13 Oct: Report of findings and in-class, demonstrations of Programming Project 4 Programming Project 5: Implement a genetic algorithm
and use it to find solutions for an optimization (workload balancing, QAP,
TSP) or constraint satisfaction problem (such as SAT, cost-based abduction,
or N-queens). Search
problem examples: · Constraint satisfaction (“cyptoquote,” Sudoku, “Jumble,” and nonthreatening queen placement), · Resource allocation (scheduling) · Optimization
|
|
October
18 |
Neural Networks Feedforward
neural networks and the backpropagation
learning algorithm |
Preparation for
Programming Project 6: Implement
a simulation of a feedforward neural network that uses backpropagation to
learn XOR. Plot the network output using at least 121 input pairs that
uniformly sample the unit square. Due 31 Oct: Report of findings and in-class, demonstrations of Programming Project 5 Recent
results: |
|
October
25, 27 |
Feedforward
neural networks and the backpropagation learning algorithm (cont’d) Application
examples of feedforward neural networks: ·
sunspot
forecasting, ·
river
stage forecasting, ·
ensemble
coding, and ·
auto-associative
networks for novelty detection and data compression |
Programming Project 6
(Backpropagation): Use
the backpropagation algorithm to train a network to recognize character representations on a 5x5 grid
of pixels. Use a representation of the characters +, -, \, /, X, and | on a
5x5 grid. Test the performance of the network using “noisy” representations
of the six characters. |
|
November
1, 3 |
Swarm
intelligence Swarm
Intelligence (see handout) Swarm
intelligence (continued) for resource allocation. Representations for
constraint satisfaction problems Some
ways to vary a swarm:
Adjust
the “velocity” computation |
Due 17 Nov: Report of findings and in-class, demonstrations of Programming Project 6 Recent
Results: |
|
November
8, 10 |
Introduction
to Adaptive Resonance Theory (ART) Design
and implementation of ART-2 networks |
Programming lab 6 (ART): Implement an ART-2 network
simulation and use it to train a network to recognize character representations on a 5x5 grid of pixels. Use a
representation of the characters +, -, \, /, X, and | on a 5x5 grid. Test the
performance of the network using “noisy” representations of the six
characters. Compare your findings with those obtained in Lab 6 (Backpropagation). Recent
results: Tag |
|
November
15, 17 |
Application
of ART-2 and fuzzy data representation to image processing/ Putman paper in
IJCNN |
|
|
November
22 |
Note:
Thanksgiving vacation begins after last class on 22 November 2010 |
|
|
November
29, Tuesday December
1, Thursday |
Presentation
of reports: |
Final report examples: 1.
comparing
the performance of various algorithms for a resource allocation problem (or a
similar one such as QAP, graph isomorphism, or N-Queens) 2.
a
comparative study of a variety of techniques on some problem of interest 3.
an
application of connectionist AI strategies to a problem of interest 4.
other
studies may be acceptable with the prior approval of the professor Reports 2011: Please
cc tagliarinig@uncw.edu with ALL
comments. |
|
December
6, Tuesday |
Presentation
of reports(continued) Last
class meeting |
Last day to submit any work, other
than the Final Examination, for grade consideration. |
|
December
8, Thursday |
Reading
Day |
|
|
Final Exam December 15, 2011 Thursday |
Test Review
|
EPeloquin’s historical character notes (verify and
compare with your own study and then use as you may see fit) |
|
END OF COURSE for Fall 2011 |
|
|
|
Notes for CSC 517: Biologically Inspired Computing |
Problem solving steps (adapted from G. Polya, How to Solve It,
Princeton University Press, Expanded Princeton Science Library Edition,
2004): 1.
What is the problem?
|
Notes for additional study |
|
DRAFT NOTES FOLLOW |
DRAFT NOTES FOLLOW |
DRAFT NOTES FOLLOW |
|
August
25 |
4.
|
Lab
0.1: Write a program that will generate all permutations of n symbols and use
the program to find solutions for the four key words in a “Jumble” puzzle.
(An in-class demonstration is due 27 August 2009). Note:
Unscrambling the key words in a “Jumble” puzzle only requires permutations of
5 and 6-letter words. Extensions:
5.
(This extension is expected of graduate students choosing
Lab 0.1)
Use your permutation generator and the letters derived from the four key
words to unscramble the response to the clue. Recall that the response may
contain multiple words. What differences, if any, does it make to simply
re-arrange all of the derived letters versus choosing a subset of letters the
size of each word and then permuting them as needed?
Alternative
Lab 0.2: Write a program to generate, count, and tabulate all solutions for
the nonthreatening queen placement problem for nxn
board sizes, where n>3. Plot a graph of numbers of distinguishable
solutions versus board size. How do your findings compare with those reported
online? Do you see a pattern? Can you formulate a rule to relate board size
(n) to the number of solutions? Extensions: 1.
Find
solutions to nonthreatening rook placement.
Alternative
Lab 0.3: Write a program to generate solutions for square Sudoku puzzles up to
size 16x16. Alternative
Lab 0.4: Review the SAT Web
site and develop an implementation of at least two methods for finding
solutions for satisfiability problems Alternative
Lab 0.n (requires instructor approval): Find a problem of interest whose
solutions are difficult to find and develop at least two distinct methods to
find solutions for that problem. Which method is better? What qualities of
the method make it preferable? Are there circumstances under which the
alternative is better? |
|
August
27 |
Resource
allocation methods, results and solution space characterization |
|
|
September
1 |
|
Sample
optimization problem: Programming report
example 1: This link connects to an example showing a study
of a resource allocation problem similar to the TSP given above. Use it to
obtain guidelines for describing the solution space for the problem you are
solving. Exhaustive
search results: 12
process results spreadsheet
(includes graph) 14
process results
without histogram data (text) TSP
by: EPeloquin: Whole space,
histogram1,
random
solutions, random
and search on one ALewis:
AllPerms, RandomResults, Report
CTucker: ExhaustiveSearch |
|
September
3 |
|
Programming lab 1: For the TSP problem,
implement a greedy algorithm, exhaustive searching and random sampling and
then compare the results achieved by these three methods. Results
(Code
samples are for illustrative purposes. Use your own discretion when
determining when, how, or if to use any such implementation details.): EPeloquin: CodeSample1,
CodeSample2 |
|
November
5 |
Introduction
to Adaptive Resonance Theory (ART) Design
and implementation of ART-2 networks |
Programming lab 5: Implement an ART-2 network
simulation and use it to train a network to recognize character representations on a 5x5 grid of pixels. Use a
representation of the characters +, -, \, /, X, and | on a 5x5 grid. Test the
performance of the network using “noisy” representations of the six
characters. Compare your findings with those obtained in Lab 3. Recent
results: Tag |
|
November
10 |
Application
of ART-2 and fuzzy data representation to image processing/ Putman paper in
IJCNN 99 |
|
|
November
12 |
Hopfield
networks and the “k-of-n” design rule for constraint satisfaction (Set
selection and N-queens) |
|
|
November
17 |
Hopfield
networks continued |
Programming lab 6: Implement a simulation of
a Hopfield network and use it at least 50 times to find solutions to the
resource allocation problem. Compare the distribution of solutions with those
obtained by the other approaches studied. What are the relative merits? Resource: Tag’s Hopfield Network Code Sample for
“k-of-n”, permutation matrices, and resource allocation examples Results: |