I wrote my first grant proposal the other day, for funding for next year. It would be a pretty big deal to get it, not because it’s super prestigious or anything (it isn’t) but because it would mean that instead of essentially working three jobs, I’d only have to work 1.5 jobs.
Here it is, for your edification. An interesting nugget is that this handful of words took a god-awful amount of time to write. I could have written five stories in the same amount of time. Hopefully that will get easier with experience.
Why do we like the things we like? We prefer certain paintings to others; music and literary critics expound at length on what makes work X superior to work Y, often against popular opinion; and all of us have had the pleasant frisson of finally figuring out where the oddly-shaped puzzle piece fits. We have strong affective feelings about a wide range of stimuli, for no obvious reason. Why is this?
It’s not surprising that people have been trying to formulate a rationale for aesthetic preferences since Aristotle. Sometimes they agree on general principles, often they do not; and the candidate explanations are either topic-specific or else too vague to operationalize. But what if a single quantifiable mechanism undergirded these disparate pleasurable phenomena, from simple aesthetic valuation (like appreciating a Magritte painting) to more abstract pleasure (like finally figuring out a calculus problem?)
They say there’s no accounting for taste, but I think they’re wrong.
The discovery of mu-opioid receptors in the parahippocampus (a principal associational ‘mixing area’ at the end of the visual pipeline), and the research that has followed that discovery, motivates a powerful idea: we get a particular chemical reward when we find patterns in the world that hit a sweet spot between predictable and chaotic. The ‘sweet spot’ comes from parahippocampal neurons that have, via competitive learning, come to code for specific patterns of visual stimuli, but whose codes have not become so ‘overfit’ through experience that spurious activity has been eliminated. Stimuli that are ‘understood’ by the population code result in an opioid rush proportional to the neural engagement, even as environmental regularities rev-down the neural (and opioid) response.
To this point the idea has been developed perceptually; but neuroanatomical and clinical evidence suggests that, far from being particular to late visual cortex, this ‘drive-to-pattern’ is fully general, and that it transcends perception and intrudes into more abstract (and purely cognitive) domains. If this is true, then we can adapt statistical coding models from perception, combine it with the mathematical formulations of information theory and knowledge-representation theories from cognitive science, and test whether generalized cognitive pleasure (CP) is subject to the same principles that characterize affect from visual scenes.
There is an extensive memory literature on serial and sequence learning that could be fit to the CP paradigm. The idea is this: a subject is presented with a stream of symbols, generated according to a context-free grammar. The subject’s task would be to discover ‘legal’ patterns from within the symbol stream, collapsing it as the patterns are induced. Each successful ‘collapse’ should reward the subject above what would happen in a control task featuring randomly-generated symbols; we predict that subjects would voluntarily spend more time on the task with symbol streams generated by inductable grammars. Further, we might expect differential results from subjects given ‘hints’ about the top-level structure of the grammar, vs. those given hints about bottom-up structure; and a host of variations on this theme.
To go beyond the rather simple sequence task described above, we need a way to generate a test corpus whose statistical characteristics can be altered to reflect patterns of differing complexity, and which can be extended into domains more sophisticated than streams of amodal symbols. Further, we need a way to quantify collected data in terms of the latent structure inherent in it. Fortunately, I have a head start on this task: last semester I implemented a machine learning system for hierarchical topic modeling using Latent Dirichlet Allocation. Adapting this LDA solution for the CP domain would allow us to generate pattern-laden stimuli in a host of modalities, from visual patterns, to musical compositions, and even text segments. We could use these stimuli as source data for the sorts of experiments described above, as well as for more ambitious experiments.
My advisors’ unique strengths will be put to use throughout this process of theoretical formulation and experimental design. For instance, the memory of chess masters for legal vs. illegal chess board configurations demonstrates that pattern recognition and memory are inextricably linked; and the CP model, which is deeply tied to pattern recognition, has much to gain from these results. Wilma Koutstaal’s encyclopedic knowledge of the memory and learning domain has already proven an invaluable resource in my literature review, and her suggestions of which results might be re-purposed to address the question of CP have been a great help even at this early stage. Further, our work on my first year project on the neural correlates of abstract vs. concrete semantics suggests several obvious next steps once the preliminary proof-of-concept work already discussed is finished.
Paul Schrater’s expertise in pattern recognition and mathematical modeling finds natural expression in the information-theoretic formulation of CP, and his current work on ‘aspiration’ uses some of the same statistical modeling techniques I propose to use for data generation and analysis. Much of Paul’s work involves how perception of regularity in the environment alters behavior, and insight from those results will inform my experimental design from a computational standpoint much as Wilma’s insight will inform it from a memory perspective. Further, Paul’s knowledge of both computer and biological vision will be critical in extending the symbolic presentation protocols of the first experiments into the image domain.
My wish-list for experiments past those described here are too numerous to list. The chief power of the CP idea is that results from disparate areas that can be stated in information-theoretic terms are suddenly amenable to analysis. Which means CP could bridge, experimentally, theories of higher-order cognition, perception, and aesthetics that have thus far been considered distinct islands of inquiry.