Thesis Committee Meeting

1
Thesis Committee Meeting

• Co-supervisor Edie Rasmussen Director
  SLAIS
• Co-supervisor Richard Rosenberg Prof
  Em CS
• David Poole Prof CS
• Tamara Munzner Asst Prof CS

20 April 2006
2
goals of our meeting

• During the meeting ---
• Review work done
• Decide on work necessary for completion
• Following the meeting ---
• Draw up a document summarizing the plan, to be
  approved by all members.

3
time line

• According to FoGS ---
• Two committee meetings
• First today to decide on the plan
• Second in December, to review work done and
  approve write up
• Degree completed in Spring 2007

4
FoGS suggestion

• Given that It was Mr. Huggetts belief that the
  work done in his first years of study would
  constitute in some way a portion of his doctoral
  thesis.
• 1/3 of thesis Code development (MemoPlex)
• 1/3 of thesis The large paper (Principia)
• 1/3 to be determined in this meeting
• Thesis should be written in manuscript style.

5
The Manuscript Thesis

• Purpose
• To gain writing experience in a format used by
  researchers in a field of study.
• To ensure timely publication.
• Format
• Constructed around one or more related
  manuscripts previously published or being
  prepared for publication.
• An introductory chapter sets the context for the
  work.
• The concluding chapter relates the manuscript
  chapters to each other, and proposes directions
  for future research.

6
(No Transcript)
7
The Thesis Plan

• in three parts

8
TOC

• Part 1 Overview
• Area, motivation, and background
• Part 2 Work Done
• Items of significant work
• Part 3 To Do
• Possible next steps

9
Part 1

• Overview

10
what I do

• Information management using associative
  spreading-activation networks
• Why
• It offers faster and more intuitive access to
  relevant information
• Why you should care
• It could improve how you access personal
  information

11
dynamic associative networks

• No hierarchies, and link lengths simply indicate
  relatedness.

Collins and Loftus, 1975
12
P-MAK the general framework

• Principles of Mnemonic Associative Knowledge
• mnemonic Relating to or intended to assist the
  memory
• knowledge Information organized by human goals
• A set of principles describing how knowledge is
  constructed
• Knowledge only makes sense wrt human
  characteristics
• If mechanized, also depends on properties of
  machines
• Knowledge depends on relations between objects

13
what is an object ?

• An object is a discrete piece of meaningful
  information, such as a document, image, or piece
  of music.
• An object is defined by discrete attributes
• An object has an activation level reflecting its
  utility.
• Objects may be linked by various types of
  relation
• P-MAK focuses on the document domain, where an
  object could be a book, chapter, passage, or
  article.
• In documents, the object attributes are keywords.

14
P-MAKs 3 types of relations

• Semantic
• Objects are linked if they share attributes
• Co-usage
• Objects are linked if they co-occur consistently
• Situational
• Objects are indexed by the time and circumstances
  in which they occur
• Other ...

15
objects relations network

• Retrieval from long-term memory, using spreading
  activation along weighted links.

unit object element attribute
Anderson, 1983
16
P-MAKs semantic network

• For retrieving similar objects
• Two key processes
• Defining objects are identified and put in
  nodes, and a classifier is used to extract
  descriptive attributes
• Relating nodes are linked if they share
  keywords. More shared keywords make stronger
  links
• Semantic links are static
• Semantic knowledge is cumulative and permanent

17
P-MAKs co-usage network

• For retrieving objects that are typically used
  together
• Assumed that there is some invisible (semantic)
  relation between them
• Persistence links are dynamic
• Links grow stronger the more objects are used
  together
• Relations that are not stimulated fade and are
  forgotten

18
forgetting is a good thing

• Forgetting is vital in a dynamic system, for ---
• Preventing available information from becoming
  overwhelming
• Focusing attention on items that have proven
  importance
• Allowing thematic drift, to stay up to date

19
P-MAKs situational network

• For retrieving objects that relate to particular
  times or circumstances
• Serves a predictive function by retrieving
  objects when appropriate cues re-occur
• Few computational models of such episodic memory
  exist (e.g. Laird, Miikkulaenen)
• Relations strengthen or fade depending on support
• P-MAK defines ---
• Temporal indexing for temporal events (w/ timers)
  
• Environmental indexing for spatial events (w/
  sensors)

20
Temporal Indexing

• Answers
• What usually happens at time t ?
• When is event e likely to occur?
• As events are observed, they are linked to index
  nodes that contain a conjunction of temporal
  units.
• Index nodes represent temporal patterns.
• The more common a pattern, the greater its link
  and node weights.

21
Temporal Indexing

• Events e0 and e1 are indexed by temporal nodes,
  and activate objects n0 and n1.

40
22
Environmental Indexing

• Answers
• Under what conditions does event e occur?
• What events are associated with sensor s?

23
project goals

• To build information management systems
• and knowledge structures that are ---
• Simple
• Efficient
• Extensible
• Inspectable
• Human-centred

24
Links to cognitive science

• Discrete objects
• Symbolic networks
• Spread of activation
• Co-usage nets Hebbian learning
• Situational nets episodic memory

25
Why model an IM system on human memory?

• Human memory is clearly effective at representing
  the statistical regularities of the environment.
• It is much studied and well understood.
• It provides users with a familiar mental model.
• Such systems could operate as cognitive
  prostheses by extending human memory.

26
Why use networks for knowledge representation?

• they can be built and edited ad hoc (cf.
  structured DBs)
• they are ideal for sparse domains (esp. semantic)
• they are easily depicted (cf. vector methods)
• they allow graph-theoretic analyses (clustering,
  arities, small worlds, etc.)
• Associative networks in particular are good for
  ---
• a human-readable representation (cf. PDP, LSA)
• finding related items quickly
• searching through browsing (navigation)

27
Prior art

• A short list of work in this area
• The Memory Extender (Jones 86)
• A Spreading Activation Model for IR (Preece, 81)
• On the Use of Spreading Activation Methods in
  Automatic Information Retrieval (Salton Buckley
  88)
• IR by Constrained Spreading Activation in SemNets
  (Cohen Kjedlsen 87)
• But none is a good fit.
• Using assoc-nets for information retrieval is
  rare --- but not because it has been
  proven ineffective it just doesnt
  seem to have caught on, compared with neural
  networks.

28
Part 2

• Work Done

29
Two main pieces of work

• Principia A paper that describes the basis of
• Information management using associative
  spreading-activation networks
• All project products can be described wrt
  Principia
• MemoPlex An implementation of an associative
  information management system, based on a
  specification written in (Hoos, 2001).

30
(1) MemoPlex an IMS

• Based on an unpublished white paper (Hoos, 2001)
• Information management system using spread of
  activation for information retrieval
• Items are linked through an associative network
• One link type, initially weighted for similarity
• Node and link strengths decay if not used

31
The Plex system

• Starting with a large corpus of documents ---
• Documents are represented by nodes.
• A classifier (here, tf-idf) is used to extract
  keywords for each document.
• Documents are linked if they share keywords.
• Link strength depends on the number of shared
  keys.
• Produces a multi-dimensional semantic network,
  where nodes (documents) are most strongly linked
  to their most similar peers.

32

• code demo

33
MemoPlex evaluation

• Pro ---
• Code base is robust and incrementally improved
• Provides essential network utilities and GUI
  components for further study
• Can be used as a diagnostic tool
• Con ---
• Theoretical basis is weak
• Semantic co-usage relations are entangled
• Interface is confusing
• Many irrelevant features (e.g. is also a Web
  applet)
• Untested

34
...but from this starting point...

• a succession of products

35
(2) AutoPlex

• The first application of spreading activation to
  spatio-temporal (ST) problems
• Implements and tests ST indexing in the context
  of automobile-driving behaviour.
• Reads ST data from automotive GPS logs.
• Introduces the idea of ST node aggregation.

36

• code demo

37
The Plan

• Write up AutoPlex as a systems paper, likely to
  be submitted to a conference.

38
(3) The Tempora ST paper

• Formalizes and extends lessons learned in the
  AutoPlex project.
• Introduces the Temporal Subsumption Graph (TSG).
• Formalizes the process of node aggregation.

39
Temporal Subsumption Graph (TSG)

• The TSG defines a hierarchy of units some are
  aggregable into longer periods (e.g. the defns
  of morning and spring).

40
Temporal Aggregation (Perfect)

• When an event is evenly represented by an
  aggregable level of the TSG.

As Friday is added
Weekday is substituted
41
Temporal Aggregation (Partial)

• In the absence of perfect support, an event is
  represented by a wff.

Given consistently even support
a wff is substituted
42
The Plan

• Perform experiments on TSGs and aggregation,
  using
• synthesized event series
• user event logs
• Add results from experiments to the paper.

43
(4) The Search Browse study

• In collaboration w/ Joel Lanir
• Examines whether a similarity network can aid
  retrieval tasks in large document corpora.
• Corpora of 2000 items consist of recent NYT
  articles, and Reuters articles from 1987.
• Uses Google Desktop for searching, and the Plex
  engine for building and navigating the similarity
  networks.

44
The Plan

• Experiment recently finished on 24 subjects (6
  per cell).
• Log parser is written.
• Initial indications suggest that there is a
  positive effect.
• Analysis to begin shortly.
• To be written up as a conference paper.

45
Part 3

• Work To Do

46
How it all fits together
test products
papers
Principia (journal)
P-MAK
sb experiment
conference paper
semantic
MemoPlex application
from Hoos, 2001
co-usage
Temporal (conference)
situational
AutoPlex application
conference paper
47
Possible next steps

• Principia
• review submit to journal Minds and Machines
• AutoPlex
• write up for systems conference or journal
• Tempora
• perform benchmark tests
• write up for conference
• SB study
• perform analysis
• write up for conference

48
CONCLUSION
49
The benefits of P-MAK

• Unsupervised, real-time learning and unlearning.
• Improves and maintains accuracy over time.
• Captures statistical regularities in the
  environment, as well as one-time events.
• Can be used to flag outlier events.
• Can be used to mimic and support human memory
  reliably.

50
Applications include

• User modeling
• Recommender systems
• Social filtering (e.g. in intranets and
  libraries)
• Situational awareness decision support
• Robotic episodic memory
• Memory prosthesis (esp. wrt elder care)

51
In sum where we stand

• A principled approach
• A reusable code base
• A wide area of theory and application
• 1 journal paper near submission
• 3 conference papers in the works

52
Thank you