Understanding connectionism December 5, 2006Posted by Johan in Connectionism, Neural Networks, Raves.
Like many Psychologists, equations make me a little uneasy. So connectionist models are typically not my choice of reading, when I have a choice in the matter. This is a shame really, because these models tend to be extremely powerful: with a relatively simple model, complex phenomena can be explained in a framework that maps closely on to what is known about how the brain works.
Take McClelland’s model of memory*. Reading the original article is painful, yet, the model itself is not complicated, when presented in a different way. The Belgian Cognitive Science Research Unit has an extremely enlightening demonstration of the model, called a Tribute to Interactive Activation. Just click individual nodes of the model (to simulate activation), and see the associated nodes light up as well. Start clicking a few nodes (e.g., 40’s, married, college), and soon, one of the name nodes receives more total activation than the others. This is a pretty neat way of representating how the memory of one concept is evoked by the activation of related concepts.
This ties in with something I noticed during this term’s Perception module: a lot of these concepts are just not well explained in the traditional, academic journal style of writing. Reading something like van de Grind et al’s (2004) gains-equalisation model is quite painful. I could not grasp what was going on until I started drawing figures for myself… And I would most likely never have bothered with going that far unless the paper had carried so much promise in explaining storage of the motion after-effect, which was the topic of an assessed practical. Similar measures were necessary to process the wealth of research reviewed in Livingstone and Hubel’s (1988) paper on the perceptual properties of the magno- and parvocellular layers of the Lateral Geniculate Nucleus. No real structure emerges until you make a table and start listing the things that the article claims the different cells can and cannot do.
Some kinds of information are better conveyed through figures, animations; anything but sentences. It is a little curious that despite this, so many textbooks and articles rely almost exclusively on sheer text. Probably, the maths that so unsettle me have similarly enlightening effects on those who speak its language. Perhaps there is some truth in my girlfriend, the Biologist’s repeated, pointed remark that “you can’t be a proper scientist without knowing calculus.”
* McClelland, J.L. (1981). Retrieving general and specific information from stored knowledge of specifics. Proceedings of the Third Annual Meeting of the Cognitive Science Society, 170-172.