Q270 – Interactive Activation Model Lab

Open http://www.psychology.nottingham.ac.uk/staff/wvh/jiam/ to load the model.

Click 'Get Lexicon' to begin.

1. Click on the Parameters button.

a. Looking at the row of default values for Feature-to-Letter parameters, what does the ratio of excitation to inhibition tell you about how feature nodes influence the activation of letter nodes?

b. In this model, what are false positive responses and how might this ratio be changed to create more false positives? When might this happen in everyday life?

c. For a word perception task using this model, which parameters feed information forward? Backward?

d. What layers can inhibit letter features? What reasonable connections (list and explain at least 2 possibilities) might be added to this list of parameters to add additional inhibition to features?

2. Click Cancel to return to the main view. The '_' character indicates an unknown letter (no features present). Run a simulation with the default parameters for 10 cycles with the word '_alt'. Afterward, click Active Nodes to get a printout of node activation over time.

a) Which letters become active for the 2, 3 and 4th letters? Why?

b) Which letters have any activation for the first letter? Which letters are more active than others? Why would this be the case if no features influence the first letter?

c) Draw a rough sketch of the pertinent nodes within the model (include short, written annotations) that demonstrates why W has much less activation than M for the first letter.

d) Switch to the graph of word activations. What word nodes have the most activation after 10 cycles? Why might that be?

3. The '*' character is defined in this model to be an ambiguous set of features between an R and a K.

a) Sketch how this ambiguous character might look in the simple feature set used in the model (you might want to use Figure 4 from McClelland & Rumelhart (1981) as a base)

b) Set Layer to Letters and letter Position to 1 and conduct simulations with ‘*ite,’ '*ind,’ *ide' and '*ick'. The model knows the words “rite,” “kind” and “rind,” but does not have “rick” in its dictionary. Explain the disparity between K and R activations in the first letter slot across the four simulations.

c) Look at the graph generated for “*ick” over 20 cycles. What is the final activation of R for Letter position 1? Does this seem consistent with your understanding of McClelland and Rumelhart's model description? Why? What parameter value(s) would need to be added or changed for the final activation of R to be around or below 0?

d) Consider the input '*i_e'. Run the simulation for 20 cycles. Sketch and explain a small network that demonstrates why there was a difference for ‘k’ and ‘r’ activations in the first letter position.

4. The default value for word-to-word inhibition is 0.21 in this model. Consider variations of this and sketch each resulting graph:

a) When the value of word-to-word inhibition is reduced to 0.01, run '_ick' for 20 cycles and look at the graph of Words. What about the words in the two groups separates them into two distinct groups? Why are so many words partially active at the end of 20 cycles?

b) Increase word-to-word inhibition to 0.3 and run the model again. How does this change the graph? Why?

c) Increase the word-to-word inhibition to 0.7. Give an explanation of what might have happened at around 10 cycles that produces this effect.

5. In a few typed paragraphs, relate the concepts of bottom-up and top-down processing to the way this model works. How does the cascaded and interactive process of the model differ from a standard information processing model in which earlier stages, when they complete their processing, send their outputs to the next stages.