Houses of Mirrors: Deeply

blabbedharborIA et Robotique

23 févr. 2014 (il y a 3 années et 3 mois)

45 vue(s)

Houses of Mirrors: Deeply
Adaptive Designs for Machine

Deborah Duong, Michael Ross

Next for MICCE: Ontological Level

Emergence of Data Driven Ontologies
from Text

Looking for High Independence of
grouping and low variance within

In other words, the highest mutual
information, lowest entropy grouping

Social Hierarchies

At INSCOM, subsumption hierarchical
trees of roles and role relations

Entities grouped into roles

Paths grouped into role relations

isa relations:

merchant isa merchant

MICCE finds Systems

Finds systemic relations in common to similar

Common paths between roles become role

Higher levels of hierarchy have more abstract

Happen to be social systems at INSCOM

Regular Structural Equivalence in Social

Can help to find terrorist organizations


Users may browse data in terms they are
used to, at any level of generalization

Ex. The query: “terrorists bombing civilians”
can find “Joe suicide
bombing subway

Hierarchy gives AI programs a gradient, a
measure of semantic distance from every
concept to every other concept, making
the space navigable.

Concepts of Concepts

We will implement ontologies by sending
concepts through the feedback loop

Concepts will form based on similarity,
split based on variance

Concepts become more independent as
dependent concepts are merged

With iteration concepts will become more
like orthogonal bases

Greater Independence

More accurate semantic distance

Helps to minimize variance

Example: Taking cosine coefficient with
50 synonyms for a word rather than a
single concept that combines them

Calculations more accurate because we
don’t make false distinctions due to noise

Another Level of Feedback

Types of Feedback

Side: Between entity and link assignments

Lower: Between parse, word sense, ontological levels

We already have feedback between parse selection and
word sense

Parses are chosen to reinforce existing patterns of concepts

Now higher level ontological categories can feedback
into the grouping of concepts

Ex. Concept of mammal needed to split “dolphin” from “tuna”

Feedback between parse, word sense and ontological
levels for global consensus on meaning

Ontologies Problematic

MICCE will approximate most likely
(highest mutual information) ontology

BUT, analysts want their own ontologies

Different experts look at same data


Data stored in primitive entities and paths

MICCE to make semantic model on the fly
tailored to ontology of who is looking at it.

Hypothesis Driven

AND Data Driven

MICCE can flexibly take in analyst input

MICCE can align its ontology to another
with very few points of correspondence

Feedback gives MICCE advantage over
other systems that generate ontologies:

Global consensus

Ability to adapt to any amount of user input

House Of Mirrors Design Pattern

In this design pattern, every thing is defined by
everything else

In MICCE, every concept is defined by its
relation with every other concept

Houses of mirrors use self fulfilling prophecy:
they are highly seedable

If an analyst groups concepts:

Collocated paths found

These help develop analyst’s concept

More consonant concepts and paths found

RELATIVELY FEW points of correspondence needed

Nonlinguistic world: Abstractions of

In text, MICCE separates different roles in
the same person and different abstract
processes that apply to these roles

Applied to the non
linguistic world, it will
find different function in the same items,
abstracting on the different processes

These processes can be abstracted and
specify simulations

MICCES align ontologies with each

2 MICCES, one in the linguistic and the other in
the nonlinguistic realm, may be aligned through
very few points of correspondence “pointing to
ball and saying ball”

They perform collocations for each other, ie,
images of cats serving to collocate the words
“kitty” and “cat”

Where there are no points of correspondence,
both would fill in the gaps in consonance with
the other


A proposed coevolving simulation system,
which is also a house of mirrors, can be used to
perform the more complicated collocations

By adapting to the seeds of both MICCES, it
helps to fuse the data

Simulacra will accept the systems that both
MICCES extract, and translate them into a a
“language of process” , like a language of

Our approach to natural language
understanding: To Understand is to Simulate