Re: 50 years later, Marvin Minsky still doesn't get it

minsky@xxxxxxxxxxxxx wrote:

This is a good example of commonsense thinking -- and how it can go
completely wrong. (That is, at least in my view.) Consider the
opposite argument, as proposed in Section 25.4 of The Society of Mind:

Why do we have the sense that things proceed in smooth, continuous
ways? Is it because, as some mystics think, our minds are part of some
flowing stream? think it's just the opposite: our sense of constant
steady change emerges from the parts of mind that manage to insulate
themselves against the continuous flow of time! In other words, our
sense of smooth progression from one mental state to another emerges,
not from the nature of that progression itself, but from the
descriptions we use to represent it. Nothing can *seem* jerky, except
what is *represented* as jerky.

Paradoxically, our sense of continuity comes not from any genuine
perceptiveness, but from our marvelous insensitivity to most kinds of
changes. Existence seem continuous to us, not because we continually
experience what is happening in the present, but because we hold to our
memories of how things were in the recent past. Without those
short-term memories, all would seem entirely new at every instant, and
we would have no sense at all of continuity, or of existence.

One might suppose that it would be wonderful to possess a faculty of
"continual awareness." But such an affliction would be worse than
useless because, the more frequently your higher-level agencies change
their representations of reality, the harder for them to find
significance in what they sense. The power of consciousness comes not
from ceaseless change of state, but from having enough stability to
discern significant changes in your surroundings. To "notice" change
requires the ability to resist it, in order to sense what persists
through time, but one can do this only by being able to examine and
compare descriptions from the recent past. We notice change in spite
of change, and not because of it.

Our sense of constant contact with the world is not a genuine
experience; instead, it is a form of what I call the "Immanence
illusion". We have the sense of actuality when every question asked of
our visual systems is answered so swiftly that it seems as though those
answers were already there. And that's what frame-arrays provide us
with: once any frame fills its terminals, this also fills the terminals
of the other frames in its array. When every change of view engages
frames whose terminals are already filled, albeit only by default, then
sight seems instantaneous.

THE PRESENT MOMENT: This is part of why we feel that what we see is
"present" in the here and now. But it isn't really true that whenever a
real object appears before our eyes, its full description is instantly
available. Our sense of momentary mental time is flawed; our
vision-agencies begin arousing memories before their own work is fully
done. For example, when you see a horse, a preliminary recognition of
its general shape may lead some vision-agents to start evoking memories
about horses before the other vision-agents have discerned its head or
tail. Perceptions can evoke our memories so quickly that we can't
distinguish what we've seen from what we've been led to recollect.

Any comments?

One thing that comes to mind is the relationship between what you describe
and Bayesian inference (of course since I am fairly steeped in the latter,
it always comes to mind for me). A second thing that comes to mind is the
notion of "predictive information" and its relationship to saliency, or
Roger Shank's idea of "failure-driven learning". Finally, there is the
question of how metaphorical hypothetical constructs (frame arrays,
k-lines, etc.) can be made into testable hypotheses.

o "Probabilities are a dead end"

The two comments I remember are that 1) they are opaque, and 2) "they don't
help much with reflective thought. Eventually, one needs to know the
reasons why and when one ought to make assumptions that one is not yet sure
about".

Graphical probability models, e.g. Bayesian Networks, encode two kinds of
information: *qualitative* information in the structure of the graph, the
conditional independence relationships between the random variables at the
nodes of the graph, and *quantitative* information: the conditional
distribution over the states of a node, given the states of adjacent nodes.
The equilibrium distribution at a node is the result both of "diagnostic
support", the lower level details from child nodes (e.g. bits of leg, seat,
and back, diagnostic of a chair) and "causal support", predictions from
parent nodes (e.g. chairs around a table are to be expected in a dining
room). Graphical probability models support causal reasoning, "recognizing,
generalizing, predicting what may happen next, and knowing what we ought to
try when expectations aren't met".

Some of these ideas are debated in journal review articles available here:

"Causality: Models, Reasoning and Inference", Judea Pearl, 2000, Cambridge
University Press.
http://bayes.cs.ucla.edu/BOOK-2K/index.html

24.2 Frames of Mind

"A *frame* is a sort of skeleton, somewhat like an application form with
many blanks or slots to be filled. We'll call these blanks its *terminals*;
we use them as connection points to which we can attach other kinds of
information. For example, a frame that represents a "chair" might have some
terminals to represent a seat, a back, and legs..."

"Default assumptions fill our frames to represent what's typical"

Sounds like a sort of Bayesian net to me.

25.1 One Frame At A Time?

After some discussion of a figure/ground image that can be seen as faces or
a candle, and of the Necker cube, comes:

"Our vision-systems are born equipped, on each of several levels, with some
sort of "locking-in" machinery that at every moment permits each "part," at
each level, to be assigned to one and only one "whole" at the next level."

This "machinery" is clearly a hypothetical construct. What sort of
independent evidence might support it? What theoretical considerations
might predict it?

Random switching and optimal processing in the perception of ambiguous
signals. W Bialek & M DeWeese, Phys Rev Lett 74, 3077-3080 (1995).
http://www.princeton.edu/~wbialek/optimization_links.html

"In the case of motion estimation [53] there is nothing deep about the
statistical mechanics problems that we have to solve, but here we found
that in cases where stimuli have ambiguous interpretations (as in the
Necker cube) the estimation problem maps to a random field model. The
nontrivial statistical mechanics of the random field problem really does
seem to correlate with the phenomenology of multistable percepts. This is
interesting as a very clear example of how the idea of optimal performance
can generate striking and even counterintuitive predictions, in this case
predicting fluctuations in perception even when inputs are constant and the
signal to noise ratios are high."

The "estimation problem" is one of Bayesian inference. It has an optimal
solution that depends on prior experience (so the grouping and labeling
that takes place, and the rate of switching between alternatives, the
"locking in", depend on individual histories of interaction with the
environment) and leads to principled predictions that agree with
experiment.


o "We notice change in spite of change, and not because of it"

"The power of consciousness comes not from ceaseless change of state, but
from having enough stability to discern *significant* changes in your
surroundings" (emphasis added).

What makes changes significant? One obvious answer is "novelty", or
expectation violations. These are low probability events, so when they do
happen then -log2(p) is high. Conversely, highly predictable events are not
very informative. Naturaly, what is predictable and what is novel will
change with experience. This brings up the notion of "predictive
information", those features of the world that predict probable outcomes.
Only a vanishing fraction of information is predictive:

Predictability, complexity and learning. W Bialek, I Nemenman & N Tishby,
Neural Comp 13, 2409-2463 (2001).
http://www.princeton.edu/~wbialek/learning_links.html

Predictive information is significant.


o "Nothing can *seem* jerky, except what is *represented* as jerky."

Several pertinent and interesting-looking papers are available here:

http://nivea.psycho.univ-paris5.fr/

But I have not had time yet to read them.

-- Michael

.

Relevant Pages

  • UV - College Fantasy
    ... College Fantasy Ultimate is now up and running on the UV website. ... Ignore the write up (for some reason I can't get it to update) and ... choose from any of the players competing in the open division. ...
    (rec.sport.disc)
  • Re: SR, frames, light and spatial separation
    ... And come back with a coherent definition for inertial frame. ... comparisson of time dilation. ... Unless you again invoke an exception to the rule..." ... So I asked you what rule you had in mind. ...
    (sci.physics.relativity)
  • Re: The amazing finish ( Phelps Gold Medal # 7 )
    ... Frame by frame, hi-res hi-speed camera. ... Just a few thoughts have crossed my mind: ... 'Hard enough to register as a touch by a swimmer ... The guy that LOST is the world record holder. ...
    (rec.motorcycles)
  • Re: jury duty in California
    ... Is it healthy for oneself or the world to not be in the frame of mind ... unhealthy for the world if people are in the frame of mind to kill ...
    (alt.usage.english)
  • Re: Again: Relativity for Thought Experiments
    ... The value is that you develop theories, make predictions, and then ... One cannot use a coordinate system whose clocks are not properly ... Relativity is about the comparison of TWO different coordinate ... one frame, SR does pertain to one frame. ...
    (sci.physics.relativity)