Intelligence by definition
From: bkaz (bkaz__at_hotmail.com)
Date: 10/27/04
- Next message: zerkanX: "Re: Bareknuckles Behaviorism"
- Previous message: bkaz: "Re: Hawkins ideas on building AI's"
- Messages sorted by: [ date ] [ thread ]
Date: 26 Oct 2004 21:20:04 -0700
Intelligence is not a content-specific ability, - we can learn
anything. Since inputs & outputs can't be assumed, the method of
intelligence is derivable only from the functional definition of
learning.
Mainstream position of artificial intelligentsia seems to be that
intelligence is a lot of different things, which I think is largely
responsible for the pathetic state of AI. Any general term refers to
different things, but they must have something in common or we
wouldn't use it to describe them all.
I define intelligence as an ability to build a predictive model of
environment by discovering patterns in past inputs & interactively
projecting them. That includes planning, which technically is a
self-prediction.
There are plenty of pattern recognition methods, but none scale
effectively with the range of inputs. I believe that's because they
don't consistently define match as a predictive potential or direct
search accordingly, & aren't selectively incremental in their
power/complexity.
Definitions:
Match is a partial identity: a sum of binary AND between uncompressed
comparands (strings of 1s),- the only way to quantify similarity
positively (though it can also be seen as a double negative: a
compression of magnitude by replacing the greater input with the
smaller).
Partial identity relative to input's magnitude is predictive: match &
diference are projected equally. Complexity theory defines similarity
as a compressibility of records, which is subjective & somewhat
arbitrary as it doesn't scale smoothly. In my approach it affects
costs, not benefits, of prediction.
A pattern is a sequence of matches between inputs (sets of variables)
& projections (last input scaled by vectors). Vectors are values that
convert the previous input into the last one by any operation:
addition, multiplication, elevation to power...
Vectors are discovered by comparison between the inputs: reverse
scaling of corresponding order: subtraction, division, extraction of a
logarithm... Comparison is made over the distance between the compared
inputs, which is a second type of vector: the one scaling inputs'
coordinates.
A pattern is predictive if a match of inputs is associated with that
of their coordinates. A priori such coordinates must be
Spatio-Temporal: physical causality is S-T continuous & similarity
declines with distance/delay. Then a secondary search is ordered by
any sufficiently predictive variable type.
Pattern's Predictive_Value is its projected cummulative match to
future inputs. Its' Net_Value = Predictive_Value - Average_Value(
produced at the same cost of memory + operations in the past).
Patterns of negative Net_V are deleted, thereby continually increasing
predictive power of a system.
Scaling:
I suggest that patterns(inputs+vectors) should hierarchically scale
their range of search & syntactic complexity as long as such scaling
increases their Net_Value, which depends on algorithmic complexity of
the input stream (lower-level patterns).
Scaling must start from the simplest possible inputs: sensory, defined
by the minimum of 2 variables: Value & Coordinate, indicating
respectively content & target area of projection. They're processed by
the simplest possible method: run-length encoding, but modified to be
scalable:
Example: grey-scale video processing:
Pixels adjacent in a scan line are compared to form 1D patterns,
adding a set of new variables for every old one: Brightness forms
Power of comparison, average Match, & cumulative Vector; Coordinate
forms Period of comparison, Number of matching inputs, & cumulative
Distance between them.
These new variables are necessary to evaluate:M*N, & to project:
V*Power *D*Period, the pattern for expanded search, increasing range
of Coordinate & possibly of Brightness. Such syntactic expansion
repeats with each new level of search for each surviving (predictive)
variable of an input pattern.
Vectors from sequentially successful comparisons are also compared,
forming higher-derivation sub-patterns, each with the same set of new
evaluation/projection variables. The resulting pattern potentially
forms a hierarchy of derivation, with multiple sub-patterns (offsets)
on each level.
Subsequently patterns are compared/integrated over increasing
dimensionality/distance: vertically adjacent 1D Ps are compared to
form 2D Ps, compared to form 3D Ps, compared to form 4D(dynamic) Ps,
compared over hierarchically increasing distance (no longer adjacent)
to form discontinuous patterns.
Thus, memory forms a search hierarchy, each level composed of
full-search units with a fixed range. Units' range expands with
elevation, up to a single top level global search unit. Higher levels'
range & syntax increase cost, so only patterns of corresonding value
are elevated to search them.
Patterns are deleted if the unit's sources shift beyond their range of
effective prediction.
Given sufficient hardware, the top level patterns will eventually
achieve & surpass complexity & generality of natural language
concepts, which means that the system will surpass the predictive
power (effective intelligence) of human mind.
Notes & Recapitulation:
This approach may seem primitive, but that's precisely the point, all
the complexity must be learnable/forgetable, otherwise it becomes a
bias.
Generalization is a reduction, & intelligent system that starts with
higher-level data (especially natural-language concepts) can't
recreate the semantic context, that is generalized real-world
4-dimensional patterns that most words in the language stand for.
Most of the effort in compression now seems to be in developing
methods for symbolic data, which I think is unfortunate. There is only
so much that one can compress out of symbols without knowing what they
stand for. Real compression must be selectively lossy, & a proper
selection is only possible if we can assess value of the data within
its semantic context.
Scalability is of the essence, & a truly scalable method must start
from minimal input complexity: the limit of resolution of raw sensory
data, from which all higher data types are ultimately derived.
Power/cost of comparison should be proportional to inputs' accumulated
match, which is initially 0.
Thus, the power must also start from the very minimum: arithmetic
subtraction between adjacent inputs.
I don't believe in combining different methods because cognition deals
with the unknown, - we can't a priori split it into different areas,
except to the extent that they're sensor/hardware specific, or levels,
except that patterns' syntactic complexity will increase with
sequential search expansion.
Any methodological differentiation must be determined by the success
of lower-level methods, encoded in patterns produced by them. The
'types' of inputs are ultimately the types of empirical objects they
represent, an intelligence should be able to learn them on its own.
Would appreciate any comments.
Boris.
- Next message: zerkanX: "Re: Bareknuckles Behaviorism"
- Previous message: bkaz: "Re: Hawkins ideas on building AI's"
- Messages sorted by: [ date ] [ thread ]
Relevant Pages
|
|
