Some time ago, a professor at a British university once told me that the introduction of yearly 50 pound "top-up" fees would corrupt education. He reasoned that if students could not completely concentrate on their work without undue influence, e.g. worrying about making money to pay for their education, how could they possibly engaged in the unbiased learning experience of the university? To American ears this sounds ridiculous. Some students accumulate hundreds of thousands of dollars in debt, and this professor is worried about his students paying 50 pounds a year!
I found the statement completely histrionic back then, but I'm starting to sympathize with him more and more these days. This has become especially acute since I left the holy order of philosophy for the decidedly greener pastures of computational neuroscience. Green seems pretty good to a former philosopher, but the closer I get to it the more I worry at times (as I am certainly wont to do!). As with those 50 pounds and the undergrads, could the mighty green seriously warp the priorities of researchers?
Like many fields in their formative years, computational neuroscience is rife with confusion: what do we study, how do we study it, who studies it? Some think that computational neuroscience exclusively refers to modelers who use precisely detailed compartmental models of single neurons, while others think it refers only to neural network researchers. A close analog in neurobiology is the difference between collecters of single cell recordings and fMRI data. That is to say, as a field, we lack a nice unifying paradigm that dictates which level of granularity we need to characterize. This leads to serious sociological (who does neuroscience?) and methodological (what is the unit of neuroscience) issues. Such questions require a lot of hard thought and experimentation, with a lot of grad students, research assistants, post docs, and professors grappling at them for years, if not decades. And, the lack of a paradigm leaves the field especially open to suggestion.
This leaves some daylight for private industry. With a few journalists noting the increasing influence of corporate interests, especially in the biomedical sciences, I wonder exactly how someone's research in the computational end of things might be affected. Ideally, as computational neuroscientists, we should use computational tools to aid the investigation of the brain. This could involve a laminar model of lower visual cortex or characterizing the mapping between retina and V1. Both are mathematical descriptions of brain function, on one level of granularity or another, but most importantly they follow the data as slavishly as possible. Above all else, tracking data should be a concern for computational neuroscientists, as in computational biology or mathematical physics--no surprise there. Still, a number of researchers have steered their research toward scaling up, i.e. leveraging heavy-duty computer power to do lots of number crunching.
I would normally welcome this scaling up with open computational arms, yet the recent advent of CUDA and NVIDIA's monetary involvement in academic settings presents an interesting marriage of academic and private interests. The power of GPGPU coding allows for massive speedup of computation, making large-scale simulations viable at a cheap price. This is by all measures a good thing, as long as scaling up one's project is warranted in the first place. If NVIDIA buys CUDA-specialized rigs for certain departments, introduces CUDA programming in coursework, and incentivizes a good deal of the process with free video cards and the like, I wouldn't be surprised by if we see a new surge in large scale simulations, which would redirect research for all the wrong reasons ("Check out how many neurons I got to run on this thing!"). When corporations come trotting solutions in search of a special kind of problem, well, expect a new surge in that special kind of problem. And, on a more fundamental level, some might just fall in love with the new toys and forget about what they were supposed to do in the first place ("Check out how many neurons I got to run on this thing!").
Then again, my department has debated building a GPU-centric cluster for use with CUDA. I am all for it, since a number of our projects are already large-scale and could benefit immensely from the substantial speedup in computation. My own work in object recognition needs speedup like mana from heaven. I could throw more images at my model and spend more time exploring parameters and equations, as opposed to watching simulations finish. Moreover, NVIDIA has no stake in our use of their products, nor do most of us plan on exploiting this to impress someone at NVIDIA with future employment in mind.
If anything, this will be a fascinating object lesson to watch unfold.
(Image from Flickr user Daniel Wehner)
