Base level activation based reasoning

1
Base level activation based reasoning
Progresses in WP4 by WICI
Ning Zhong, Yulin Qin, Shengfu Lu, Jia Hu, Haiyan
Zhou, Erzhong Zhou
The International WIC Institute/BJUT
2
ACT-R Meets GrR

• ACT-R, a computational cognitive architecture
  based on an information processing modeling of
  human brains, is used to analyze our studies and
  develop a plugin
• The idea of chunk and information organization
  structure (multi-level granularity structure)
• Spreading activation theory and information
  retrieval/selection strategy

3
Activation in ACT-R meets GrR
frequency recency
situation, e.g. task demand
personality, knowledge system, e.g. semantic
context
4
Index Selection List

• Selection based on scalable granules
• Off-line index selection sample
• On-line sample reasoning

5
Basic-level Based Index List

• Simplified equation
• Based on frequency and recency
• Need not to compute and update Bi every time(day,
  minute, etc.)
• Demo hotel selection

decay parameter
6
Demo Hotel Selection
frequency
7
Demo Hotel Selection
recency
8

9
(No Transcript)
10
(No Transcript)
11
2. Human Heuristic Search in Problem Solving
Progresses in WP4 by WICI
Yulin Qin, Ning Zhong, Haiyan Zhou, Shengfu Lu,
Jie Xiang
The International WIC Institute/BJUT
12
Contents

• Motivation and Goal
• Methodology
• Experiments tentative results
• Discussion Links to LarKC
• Future work

13
Motivation

• LarKC focuses on the huge scale requirement in
  the Web and other IT areas caused by the
  explosion of information and knowledge (e.g.,
  reasoning about 10 billion RDF triples in less
  than 100 ms). (Fensel and Harmelen, 2007)
• One of the central ideas to overcome this problem
  proposed by LarKC is limited rationality,
  including heuristic reasoning. (Fensel et al.,
  2008)
• Human beings have developed sophisticated
  heuristic search skills in reasoning, problem
  solving and decision making.

Can we be enlightened by human being?
14
Motivation

• heuristic reasoning in the web

Goal

• Neural basis of human heuristic problem solving
• 2. Hints on improving the performance of
  searching and reasoning in LarKC.

15
Methodology
fMRI

• fMRI experiments
• fMRI data analysis
• Confirmation analysis, in which the major ROIs
  (Regions of Interests) were found by ACT-R group.
• Exploratory analysis, in which the ROIs will be
  defined based on the activations.
• Multi-Voxel Pattern Analysis, Extracting the
  activation patterns of each voxel .
• Modeling
• An ACT-R model will be built to
  predict/simulate the behavior data and the BOLD
  effect of fMRI in selected ROIs.

16
Paradigm Simplified 4 x 4 Sudoku
4 x 4 Sudoku

• It is simple (good for fMRI)
• It is still a problem (involving all aspects of
  problem solving)

Start State
Goal State
17
Heuristic Rules in 4 4 Sudoku
Simple
Complex
The heuristic rules are simple, clear and easy to
use.
18

• How to apply a heuristic rule?
• How to find an appropriate heuristic rule?
• gt Simplified 4x4 Sudoku
• (Today)
• How to learn new heuristic rules?
• (Future)

19
Study 1 The Retrieval and Application of
Heuristic Rules

• 22 event-related parametric designed fMRI
  experiment
• Steps needed to solve the problem, 1-step vs.
  2-steps.
• Difficulty to solve the problem, simple vs.
  complex.

a. 1-step simple b. 1-step complex c.
2-step simple d. 2-step complex
20
Study 1 Experiment Procedure

• Training
• Training participants to get familiar with the
  heuristic rules
• Scanning
• Participants were encouraged to finish the
  problem as correctly and quickly as possible.
• Both behavioral data and BOLD signals were
  recorded and analyzed.

The procedure of a trial.
21
Study 1 Results Behavior Data
22
Study 1 Results fMRI Exploratory Analysis
23
Study 1 Results fMRI Multi-Voxel Pattern
Analysis

• Extracting the activation patterns of each voxel
  to get more information to explore the mechanism
  of applying heuristic rules.

24
Clustering Activation Pattern

• Brain areas of different activation patterns

25
Study 2 The Selection of Heuristic

• 22 event-related parametric designed fMRI
  experiment
• Task type, with question mark vs. without
  question mark (anchored or non-anchored).
• Difficulty to solve the problem, simple vs.
  complex.

2
?
3
1
2
3
1
?
2
4
2
3
2
3
4
1
4
1
4
2
a. Anchored simple

• Anchored
• complex

c. Non-anchored simple
d. Non-anchored complex
26
Study 2 Experiment procedure

• Training
• Training participants to get familiar with data
  glove
• Training participants to get familiar with the
  heuristic rules
• Training participants to get familiar with
  responding system
• Scanning

Time
Max 20
2
s
2
s
s
2
s
10
s
2
2
3
3
4


1
1
1
1
Response
Stimulus
Feedback
Trail Start
Rest (ITI)
27
Study 2 Results Behavior Data
28
Discussion and Links to LarKC

• Study 1 suggested that using heuristic rules,
  even very simple, might involve the activation of
  vast brain areas related to goal control, problem
  state representation, memory retrieval and so on.
  
• gt Heuristic search seems involving complex
  information processing processes that requires
  good organization of the heuristics and the
  appropriate way to use heuristics.
• Study 2 showed that the mean Reaction Time and
  Accuracy of anchored task were better than that
  of no-anchored task.
• gt It seems that cues provided before reasoning
  can improve reasoning performance, such as
  reasoning speed and accuracy
• Only tentative results, needs much more work.

29
Future Work

• Modeling heuristic problem solving based on ACT-R
  to reveal the dynamic characteristics of the
  information processing in heuristic problem
  solving.
• How to form new heuristic rules?
• Hints to LarKC ?

30
Thank You !