[FRIAM] Predictive coding basedon deep learning

[glen∈ℂ]

It's in these crevices of the discussion that it becomes obvious that the "painted surface" analogy [†] fails completely. This is why Hoffman's "interface" theory is so attractive. The same core point is made, just with more explanatory power. It's analogous to how we play video games. Good examples are fighting games where you push buttons in different sequences to make the (physics-based) avatar do things in its environment. Our mental map of the controller interface has literally no similarity to the "physical" realities inside the game. But it *works*. You can control the avatar even though the control surface is nothing like the physics it's controlling.

As Hoffman points out in one of his papers, it can even be *bad* to have an accurate understanding of the controlled system. Competent players don't get hung up on, e.g. whether a rapier or a broadsword being wielded by their avatar reflects the real thing. They simply (randomly?) try lots of game play techniques and "git gud". What's being selected for is not a good/true/real mapping, but an effective mapping.

cf. Here's a video I haven't had the chance to watch yet: https://iai.tv/video/the-reality-illusion
FWIW, I find Hoffman's style and attitude in both his writing and presentation very off-putting. But I really like the fundamental idea.

[†] Which I first heard of by D. Dennett, I think, where our mind is projecting a movie onto an opaque screen. The world is projecting an image onto the other side of the screen. And there's some functionality to the screen so that when the two projections *match*, there's some feedback to both. When the projections are too different, then perhaps there's negative feedback.

On 7/29/19 11:28 AM, Marcus Daniels wrote:
> Steve writes:
> 
> < What I'm trying to expose is the meta-heuristic of being a facile model builder/adopter/fitter... and how our technological prosthetics (precut colored plexiglass and stain-by-number patterns or GPS/routing systems that present opaque-to-the-user preferences or predictive SDE programming environments).  >
> 
> When technology doesn’t work, take it apart and figure out what is wrong with it or how it could be improved.    Human experts, or skilled practitioners, can hurt more they help because they have no incentive to unpack their expertise into reusable automated systems.   The trick is to look at skills as technology and to be facile evolving the technology.