Posts tagged with computer vision










Gerald Jay Sussman on biology & computation.

  • The human genome is 1 GB. So is Windows OS.
  • We have no idea how to program a 10¹²-unit thing like a human, or a cow.
  • A salamander regrows three elbows if you break off its arm and reattach it — responding to local errors.
  • We have no idea how to write computer vision in a few steps like neurons do.
  • Program efficiency doesn’t matter.
  • Memory is free right now.
  • Computation is free right now.
  • For a million bucks you can get a seriously computer—but what to do with it?
  • Most of the cost of a computer program is paying programmers. —Huw Evans
  • Yet we spend so much time modifying existing code. —Huw Evans
  • Correctness doesn’t matter. (Getting a reasonable answer is fine.)
  • Security doesn’t matter. (Humans are attacked by parasites all the time.)
  • Biological systems are written to solve problems that the designer didn’t foresee.
  • This LISPian stuff of writing programs that write programs looks very awesome; what I’ve pretty much always wanted to do with computers. (I tried and failed to use PHP’s multiple dollar signs to helpful effect.)
  • Some stiff-matrix stuff that you can understand if you watch some Gilbert Strang videos.
  • The point being: with bindings and such you can write a program that’s somewhat robust: performing operator overloading or similar things so that you can just tell the computer basically what you mean. (Rather than having to be so specific.)
  • "Mathematical symbols are impressionistic" — just think about how the fnof; symbol is used — yeah, some function, whatever, ya know what I mean.
  • Systems that accept a wide variety of inputs and only give a small range of outputs.
  • Your cells have about 1 GB of ROM and a few megabytes of RAM.
  • "Galileo discovered the value of a lie — to figure it out without the friction and then put it back in."
  • Propagators: Independent Stateless machines connecting stateful cells
  • Satisficing / monotonically increasing local information about a referent.
  • Synchronising problems in parallel computation go away.
  • EE, not CS, point of view. (and I daydreamed off into economic theories using this circuit-diagram thinking)
  • "I hate modern languages, even the ones I invent! … There are no names for intermediate parts [of expressions]."
  • Min 44. Maybe this is an insight into explaining statistics to programmers. Errorful processes go in; this would be like multiple inputs (maybe a distribution assumption, maybe a bound on the error, maybe an independence assumption or equal variances) — and an errorful result comes out. Just like division-of-integers has two outputs—quotient and remainder—so do statistical processes spit out a ton of outputs as well:—answer(s) like beta;’s, error estimates (p’s and t’s), and tons of different ways of looking at what might be wrong with the assumptions (Durbin-Watson, structural F statistic, White estimators). Of course you could turn these outputs into Boolean by saying like “alpha; <.05 goes to TRUE” — but really the continuous alpha score is better.
  • Expert reasoning in circuit diagrams as local reasoning between neighbouring pieces of the circuit diagram.
  • Improving measurements by making independent measurements of the same thing by different methods.
  • Tracking of provenance & shadow premises—….daydreaming: relationship to religious faith/scepticism.
  • Giggling when logic subsystems conflict.
  • Truth Maintenance Systems — the ability to back out
  • Finding lies — medical statistical results
  • Globally inconsistent worldviews that are locally consistent.
  • Aaaand, find me the consistent sub-worldviews that are consistent.

(Source: infoq.com)