Lecture 8 - Paul Quinn (Wash & Jeff College)

---

Can a Connectionist Modeling Approach Enrich our Understanding of the Development of Categorization Abilities in Human Infants?

Introduction: Percptual Categorization

"One of the fundamental tasks of the nervous system is to categorize." Edelman (1987).

Perceptual categorization refer to the ability to recognize discriminably different entitties as members of the same category based on some internalized representation of that category.
     This has the advantage of reducing physical variability, organizing storage and effiecent retrieval, and helping us cope with novel stimuli.
     Historically, Hull (1920) talks of categorization as forming from large experience with the various members of the category. On the other hand, Leach (1964) suggests that categories are in fact formed as part of a social continum (cats are those things that are labeled cats). Finally, Rosch et. al. (1976) agrees with both, arguing that there are regularities present and recognized both before and after language. Indeed, she suggest that their are three levels, superordinate, basic, and subordinate. Of these, the basic is the most fundamental. As support, Rosch, Mervis, Gray, Johnson, & Boyes-Braem (1976) found that children where great at organizing at the basic level but not so good at the superordinate level.
     Quinn has subsequently extended this research to 3-month-old infants using a Familiarization-Novelty Preference Procedure.

EG.

f1 + f2, f3 +f4, f5+f6, Then, fn + N

(Insert pictures of Fagan box and dot patterns here)

However, many reviews said, yeah but what about natural category formation? So did cats, dogs, and birds (equalized by size). (so half the babies saw a boatload of dogs followed by a dog and a bird, and half saw a bunch of cats followed by a cat and a bird). Found that 15-month-old infants prefered birds 61.65% and 63.63% over the previous category (dogs and cats respectively). This was not the result of stimulus salience, or that the cats and dogs all looked the same.
     Next, found that babies formed categories for dogs and cats as well. Hmmm, cats and dogs are pretty similiar, how could babies be doing this? Well, could be overall gestalt, or certain features, or maybe, THE FACE. So, presented three conditions, full animal, just head, or just body. Found 63.58, 67.06, and 48.14 % respectively. Notice, they couldn't make the discrimination by body alone. So, to answer headless body critics did final set which switched the heads, and infants prefered the new head. Finally, started to study the fine features of the heads to see what children were responding to.

Then, Mandler & Bauer (1988) came out, and everything was turned on its head. She notes that the original test could have been solved with basic or global discrimination ability. Indeed when tests within superordinate categories (eg. car v. truck). children can make super-ordinate distinctions (car v. cat) much sooner. But, wait! How does she explain the early perceptual findings.
     Mandler proposed a double dissociation between the perceptual levels (tapped by the preference procedure), and global conceptual views (tapped by the object exploration procedures). The problems with this however, are many.

• Cognitive Economy - Why have two different representation for each object category?
• Conceptual Bias - Self-starter vs. Non-Self-Starter. What about taxis?
• Procedure? Object Examining vs. Familiarization-Novelty Preference, are they really that different?
• Evidence? Found young infants do distinguish global categories easily.

This leaves us with a few questions:

• What is the course of category development?
• What representations are formed by newborns, 1-, and 2-month-olds?

Modeling the Data: Global to basic category learning

Created a model which took as input the various properties of the various classes (eg. head length, head width, eye separation, nose length, nose width, ear separation, ear length, etc.) and had various output categories (eg. animal, dog, furniture, cats, etc) In addition, was careful about how the features were presented to the network.
     Found (not too surprisingly) the network learned global categories first (mammals and furniture) and later distinguished the basic level categories. Even when didn't explicitly teaching the global category level. Even in a no face, no tail network (EXCEPT FOR THE CAT AND DOG DISCRIMINATION). Horray!
     Next, changed the number of hidden nodes from 8 to 10 to 11. Found that the greater the number of hidden nodes, the larger the number of hidden nodes which initially respond to the global level. However, as learning proceeded all the networks had 1 or two hidden nodes responding toglobal attributes. Thus, the bulk of the remaining hidden nodes were now coding for basic level catergories. In essence, the system (child) develops global categories first, but by adulthood has turned the bulk of the memory to hidden nodes.
     In addition, did run with autoassociator (so as to avoid teacher, training issues), and found it not only learned the same, but also seemed to follow the same developmental trajectory as acctual children.

New Data

Indeed, 2-month-olds, do learn the global task (Mammals vs. Furniture) but not the basic level task (Rabbits or Elephants). Exactly in accord with the predictions of the model.

Implications

Got a new great grant in Exeter and TLEARN RULES!