Am increasingly convinced that concepts need to be malleable, nebuluous – in Hofstadter’s words, fluid. To do this, you need to be able to tweak them in order for them to most cleanly fit into the niche that the environment has defined for them (given the way we interact with it). This is kind of what I think that Holland et al. (Induction) are getting it with their arguments for directed environmentally situated and embodied induction, though I’m putting Lakoff’s words into their mouths.
My current plan is to try and represent concepts as functions. I need to elaborate on what I mean by this (xxx). I considered going all the way back to basics, and trying to use Turing machines as the most elementary building blocks of these concept-functions. I’m now thinking about jumping up a few levels of abstraction to some high-level programming language. But then we’re back to the problem of brittleness. What about, instead of using source code concept-building blocks, using neural nets? This too needs much elaboration. (xxx)