I like to describe my techniques as "not novel, just new." I'm absolutely building on the work of many others in academia. I've spent time with Engert at Harvard, Rex Kerr at Janelia, and some of Deisseroth's students, among others (including Boyden, Ramanathan, Manuel Zimmer in Vienna, Aravi Samuel, George Church, and more). I've only briefly met Bargmann and I loved the advice she gave me: "Hmm. That's probably not going to work, but...it's really worth trying."
Unfortunately, most academic scientists are not in a position to risk their careers on such a bold and encompassing proposal. However, I benefit from this in some ways, because academics who are interested in this problem often work on a small piece of it instead of going for the whole thing, and are then incentivized to share that piece with me to integrate with all the other pieces, so they can see their contribution realize its full potential.
I wouldn't underestimate the technical demands of the project - it certainly would have been unthinkable 10 years ago. Sydney Brenner once famously wrote: "Progress in science depends on new techniques, new discoveries, and new ideas, probably in that order." However, you're right to point out that it's really the abstract methodology of delegating experimentation to a machine which truly distinguishes my technical proposal from related work. I have to admit that I haven't worked out in detail what the math will look like, though http://arxiv.org/abs/1103.5708> is a pretty good start. The tricky bit is defining the probability space (that is, the family of models under consideration). As a probability space, it must have a measurable structure. But for efficient and effective inference, it should also have additional structure, like a vector space. Yet any particular choice of vector space representation will trade off dimensionality against non-convexity, so it's desirable to have multiple representations of the same space. I've been taking a "cross that bridge when I come to it" approach, as I'm occasionally reminded by academics that if all I manage to do is collect a bunch of time series data about hundreds of neurons simultaneously, that would still probably be scientifically interesting and novel, even with ordinary, human-driven analyses.