Name: ___________________

Student No: ___________________

Faculty: ___________________

Signature: ___________________

THE UNIVERSITY OF QUEENSLAND

FINAL EXAMINATION, SEMESTER II, NOVEMBER 2001

COGS2010

LABORATORY INTRODUCTION TO MODELS IN COGNITIVE SCIENCE

Time: Two (2) hours for working

Ten (10) minutes for perusal before examination begins

General Directions to Candidates

There are twelve questions on the paper. You should attempt all questions. Each one is worth 10 marks (Total marks 120).

The examination has been set to take the full two hours, so answer the questions that you know first, and only after answering them, go back to the questions that you are less certain about.

Answer all questions on the examination paper. If you need additional space use the back of the paper or an additional exam booklet and clearly label each answer.

Question 1

marks

a) Consider an interactive activation and competition (IAC) network that represents different visits to professionals. The network has pools of features for {doctor, dentist, accountant}, {sprained ankle, toothache, regular checkup, tax return} and people’s names. Draw an IAC network that will represent the patterns for (1) Xanthe’s visit to the doctor with a sprained ankle (2) Yummi’s visit to the accountant to do her tax return and (3) Zach’s visit to the doctor for a regular checkup.

4

b) How many positive and negative weights does your network have? Explain your answer.

1

c) If Xanthe sprained her ankle again, and visited the doctor for a second time, could that be represented in the network as a separate visit? If so, explain how, if not, explain why not.

2

d) What aspects of human memory was the IAC network designed to model? Explain how it achieves those behaviours.

3

Question 2

marks

a) Cohen, Dunbar and McClelland’s model of the Stroop effect was originally developed to demonstrate that a learning model could account for the differences in response time for colour naming vs. word reading. How were response times calculated in the model?

2

b) How was does the model allocate attention either to word reading or colour naming?

2

c) The Stroop effect is also found in other tasks, such as reading numbers vs. counting them. What aspects of the original CDM Stroop network would you modify to model the differences between reading numbers vs. counting them?

4

d) In the original CDM model, word reading was much faster than colour naming. In reading numbers, the difference is not so large. How would you model the smaller difference in your design compared to CDM’s model?

2

Question 3

marks

a) What is the difference between modelling recognition and recall in terms of the output of a memory model?

2

b) Why are distributed representations typically used in modeling human memory?

3

c) What aspects of Hebbian learning (in matrix models) are useful for modeling human memory?

3

d) From a memory modeling perspective, what problems would prevent adding a hidden layer to a matrix memory model?

2

Question 4

marks

a) Draw a neural network that has one input, one hidden and one output unit (the 1-1-1 network). The hidden and output units are sigmoid units.

2

b) How many weights are there in this network (include biases if relevant)?

1

c) The task of this network is the identity mapping - simply map the input (0 or 1) to the output (0 or 1). Explain in your own words how the backpropagation learning algorithm works on this example.

5

d) After learning, would the network produce the exact outputs, 1.0 and 0.0? Explain why or why not?

1

e) Why wouldn’t the Hebbian learning rule be appropriate for training the 1-1-1 network?

1

Question 5

marks

a) Self-organizing maps (SOMs) have been used to study the organization of topographic maps for different areas in cortex. Based on the design for oriented lines in visual cortex studied in the lab, explain how a SOM network could learn to represent a simple tonotopic map given the tones (frequency bands) as input. Use a one dimensional map layer (instead of the conventional two-dimensional map). Include a diagram of your network.

6

b) What would be appropriate initial values for the weights (and biases if relevant)? Justify your answer

2

c) Would the SOM accurately represent tones of higher pitch than those it saw during training? Justify your answer.

2

Question 6

Marks

a) Define conceptual slippage (as the term is used in Copycat).

1

b) Give a real-world domain in which context-dependent conceptual slippage is required and explain why it is required.

2

c) The structure mapping engine (SME) is an analogical mapper that takes hand-coded representations of source and target domains and finds mappings based on the similarities between the given representations. How does Copycat differ from this approach when making an analogy with respect to its mapping and representation formation?

1

d) As a cognitive model, is Copycat’s approach more or less similar to how humans make analogies? Justify your answer. (marks will only be given for the justification)

1

Copycat question continued:

A possible solution derived by Copycat to the problem abc : abd, kkjjii : ? is lljjii. This solution is driven by the fact that abc is a successor group to the right and kkjjii is a successor group to the left, allowing a correspondence to be formed between them. For the solution lljjii to be formed in the Workspace:

e) what initial and modified rule would allow this to occur ?

1

initial rule: replace ________________________ by _______________________

modified rule: replace ________________________ by ________________________

f) The slippages between the concepts used in the initial and modified rule are either implied by or incorporated in the correspondence that is built between the groups abc and kkjjii. Denote what descriptions would be attached to the two groups, the slippages that occur when the two groups are mapped, and what implied slippages are required to derive the modified from the initial rule above.

1

Copycat question continued:

g) In the following diagram, fill in the missing descriptions, groups, bonds and correspondences that are consistent with your answers above.

3

k

k

j

j

i

i

Question 7

marks

1. Explain how time is incorporated into the following neural network models:

a. The Buck and Buck firefly model

b. The IAC model of the Jets and Sharks database

4

2. Explain how Nettalk maps “time into space”.

2

3. If Nettalk had recurrent connections, how could letters in context be processed without mapping time into space?

2

4. For each of the following architectures, state whether the network has recurrent connections or not:

a. The IAC Jets and Sharks model

b. The matrix memory model

c. The CDM Stroop network

d. Copycat’s slipnet

2

Question 8

marks

This question refers to Hinton and Nowlan’s simulation of the Baldwin effect.

a) Describe the initial population.

b) Explain how the fitness function rewards learning.

2

c) Explain how a simple genetic algorithm (SGA) creates successive generations.

5

d) In Hinton and Nowlan’s original simulation, the chromosome had 20 genes, the initial population was 1000, and the number of guesses was 1000. What changes in outcome would you expect if the number of guesses was decreased to 100?

1

e) In the lab class, the simulator presented the results of each run as a graph. Explain what the x- and y-axes represent; and what is plotted on the graph.

2

Question 9

Marks

a) Describe how elitism is implemented in a selection algorithm?

b) What effect does it have on tournament selection in Hinton and Nowlan’s simulation of the Baldwin effect?

2

c) What effect does mutation have on variability in an evolutionary algorithm?

d) What effect does crossover have on variability in an evolutionary algorithm?

2

e) Describe Lamarkian evolution.

f) Why is the Baldwin effect not considered Lamarkian?

2

g) In simulations of the Baldwin effect that use crossover without mutation, both roulette wheel and tournament selection occasionally have residual question marks at the end of a run (i.e., after the population has converged). These question marks are monomorphic (meaning that every individual in the population has the same allele for that gene). Monomorphic genes have different causes in roulette wheel and tournament simulations. Explain the different aspects of roulette wheel and tournament selection that result in residual question marks in some Baldwin simulations.

4

Question 10

marks

a) Define the term “evolutionarily stable strategy”

b) Explain why game theorists consider that any strategy based on pure altruism would not be an evolutionarily stable strategy.

2

c) What critical factor distinguishes the Prisoner’s Dilemma (PD) from the Iterated Prisoner’s Dilemma (IPD) problems in game theory? Explain why it makes a difference.

2

d) Give real world examples that have been considered as good examples of PD and IPD (one example of each). What conclusions for agents (societies as a whole, individual humans, animals or synthetic) that trade together can be drawn from PD and IPD?

2

e) Define the term “strategy” in game theory.

f) Illustrate the concept of “strategy” by explaining the nice (always cooperate), nasty (always defect), tit-for-tat and random strategies in the IPD.

2

g) Consider two simulations:
1. a population of 30 nice, 30 nasty and 30 tft
2. a population of 45 nice and 45 tft.
What differences would you expect in the number of nice in the population over time in the first simulation compared to the second? Explain you answer.

2

Question 11

marks

a) Simulations such as the Baldwin effect, IPD and neural network models do not – on their own - allow a researcher to explain a physical phenomena, such as the Stroop effect, or the evolution of cooperation. In general, what types of benefits for understanding real-world phenomena can be gained from simulations? Use CDM’s Stroop model as an example.

5

b) Explain why a group selection argument, such as the idea that genetic variability speeds up evolution, is not a plausible reason for the high level of sexual over asexual reproduction in mammals.

2

c) Explain the Red Queen argument for genetic mixing.

3

Question 12

marks

a) What is a unified theory of cognition, and why does Soar qualify as one?

4

b) Briefly describe the architecture of SOAR.

4

c) What are the benefits and limitations of Soar's chunking procedure?

2

End of Exam

Question	Marks /10
1
2
3
4
5
6
7
8
9
10
11
12
Total/120