Lecture 14 - Patrick Juola (Oxford)

---

Neural Network Models of the Past Tense:
Why double dissociations don't mean much.

Touchstone for connectionist models. Often degenerates into "no you can't," yes I can." Still, there is definately a phenomenon to be described. Interestingly, children seem to learn the correct irregular forms of the past tense and then forget them, only to acquire them later. Rumelhart and McClelland first modeled this apparent "rule based" performance in a network in 1980, apparently showing that connectionist nets can learn rules.
    More recently, Juola and Plunket have attempted to model past tense in a more realistically sized corpus (based on Brown's word norms.) Naturally, it learned and behaves in a similiar manner to people. However, it is in how it breaks when we lesion the network that we want to focus the talk today.

Double Dissociations

Patients with A not B, and Patients with B not A are often seen as irrefutable evidence of two seperate processing systems being involved. For example, deep vs surface dyslexia, living vs non-living semantic aphasias, patients who are good at the regular and those who are good at the irregular past tense (Wilson and Tyler).
     So for the past tense, "when memory fails rule prevails," chant the dual route modelers of inflection. Thus, neural networks can't be used in the brain.
     Of course, Plunkett and Juola would like to demonstrate that this simply isn't true. As evidence, pipe fitting structured nets. First, would like to note that box and arrow conectionist networks can be explained by the dual route, but are interesting in that they all learn in the same way, just different inputs and output pairings.
     Next, one could see a double dissociation if have an S curve of task performance. That is, if the test is linear, than really damaged patients might performed differently than mildly damaged patients, realitve to that task.
     Finally, Plunket and Juola lesioned a network model trained on a corpus of 32000 words. They lesioned the exact same network 4000 times. Found that not only did performance go down, but it went down in a bell curve like shape. Thus, sometimes the performance acctually got better, and if do a scatterplot of performance after loss find can get double dissociations in an "unstrcutured" single route system. In effect, randomness within the system can create spurrious examples of a double dissociation. Even if do another scatterplot for words in the begining of the alphabet vs words in the end of the alphabet.

Final Conclusion

We need a way to seperate artifacts from real functional seperation. As we have seen this could simply be by chance.