Lecture 2 - Peter MacLeod

---

Simulating Deep Dyslexia
by lesioning a model of reading via semantics

Outline

• Deep Dyslexia - Patients produce both semantic and visual errors in reading.
• Hinton, Plaut, Shallic Model - reccurrent model which models reading via semantics.
• Result - Both semantic and visual erros are produced bythe model.
• Attractors - model achieves its effect via use of acctractor basins.

Connectionism & Neuropsycholgy - Connectionist modeling lends itself naturally to simulation of neuropsychological deficits, because one can create a model and simply remove some connections or units AKA "Lesion" the model.
     In addition, connectionist nets have a number of very interesting properties which are similar to actual human behaviors.

• Graceful Degradation
• Resistance to noisy input
• Errors similiar to human errors.

Patients with deep dyslexia make both semantic and visual errors.

Semantic: NIGHT --> "sleep"
Visual: SCANDAL --> "sandals"

In addition, patients show Part of speech effects:

• Content words easier than function words
• Concrete words eaiser than abstract
• Pronounce non-words like SLINT

Inputs:28 position specific letter units
Hidden: 40 Hidden units
Output:68 semantic feature units.
Cleanup: 60 Cleanup units (recurrently connected with the sememe output units).

Task: To turn on the correct sub-set of semantic features for each input word. This is a real problem beacuse the relation between semanitic similarity and visual similarity is purely arbitrary. That is, semanitically related words aren't usually the same. CAT is less like DOG than COT, but dog is the semantically related.

Procedure: Present a grapheme/sememe pairing. First, let the network settle to some, and then incorporate backpropagation.

Results: In this manner, for example, although COT and CAT might enter sememe space in about the same place, the cleanup units (which encode semantic relationships) will pull the respective words into their own sememe attractor basin.

Lesioning: By removing connections, units, or adding noise, the network produces both semantic, visual, mixed, and unrelated errors. However, this is much better than just chance. First, there are virtually no unrelated errors. Second, if disrupt from grapheme to hidden, get visual (but some semantic), and if disrupt from hidden to sememe get mostly semantic errors (but still get visual errors). Thus, any lesioning at all will produce both Visual and Semantic Errors. Why? Because, lesioning has the effect of changing the basins of attraction, and those basins are the combined product of all network activity. Moreover, such change really increases the likelihood of V+ S errors. Indeed, this is true of the model and patients.

Note: Even if drastically revised the connectivity and/or move the cleanup units to the hidden layer, the model still provides the same qualitative data. What's more, if use a different learning algorithm (say Hebbian) still get the same results!!!

So what does all this tell us?

• Computational models like this provide specific, testable results. Without it most cognitive models are left with "ineffective handwaving." (I think this "box" would do this, but can't tell about the interaction.)
• One can perform experiments on these models and make predictions about patient and normal performance on tasks. Indeed, MacLeod did find that normals will make similar errors to the network. For example, network predicts that the onset of a visual mask will shift the trajectory of the setteling attractor. Namely, early onset produces visual errors and late onset produces semantic errors.
• Although some might make the criticism that this is too powerful a method and "tells us nothing because one can model anything," the fact remains that these models reveal some very peculiar behaviors which no one would have predicted otherwise, behaviors that are exhibited by humans.