This document describes a proposed description logic architecture to solve Bongard problems (BPs), a set of 100 visual puzzles designed to test human intelligence. The architecture uses semantic web technologies like ontologies, description logic, and SWRL rules to represent the puzzles and derive their solutions. It was able to solve 65 out of 100 BPs by defining the objects and their properties in an ontology, applying inference rules to detect similarities and differences between the left and right sides, and checking if the inferred concepts are consistent with the unique solution. The approach formalizes BPs in logic to leverage reasoning abilities unlike past work that used pattern matching methods and solved fewer problems.
Raman spectroscopy.pptx M Pharm, M Sc, Advanced Spectral Analysis
Public Hearing of Ph.D. Thesis of Jisha Maniamma /マニアマ ジーシャ 学位論文公聴会
1. A Description Logic Architecture
to Reconstitute the Minimal Semantic Representation Equivalent
to the Unique Solution of Bongard's Ill-Posed Problems
Considered to be Incapable Without Human Intuition
JISHA MANIAMMA
(Supervisor: Dr. Hiroaki Wagatsuma)
Department of Life Science and Systems Engineering
Graduate School of Life Science and Systems Engineering
Kyushu Institute of Technology
1
2. 2
The summary of our work
Solution- using Ontology based Knowledge Representation
(𝐿𝑖∈ 𝑃𝐴) ⋂ 𝑅𝑖 ∈ 𝑃𝐵 ⋂ (𝐿𝑖∉ 𝑃𝐵) ⋂
(𝑅𝑖∉ 𝑃𝐴) ⋂ (PA ⋂ PB= ɸ)
SA ⋂ SB= ɸ
Abilities of Human brain
Concept networks
Frames
Meta data
Filters
Sameness detector
Problem- A step towards General AI by Solving Ill-posed problem
N-features Considered Features
2006 1995 2020
3. Table of contents
1. Introduction
• Ill-posed problems, Bongard Problems (BPs), Hofstadter’s
Idea and the research objective of the dissertation
2. Semantic Web Technology
3. A RDF-based knowledge representation to solve a
typical BP
4. Proposed description logic architecture to solve 65 BPs
out of 100 BPs
5. Solving BPs with Dependent Properties
6. Discussion and Conclusion
3
4. ←→ solutions do NOT exist,
←→ the solution is NOT unique,
If it is not well-posed, it is called ill-posed problem and
meaningful in a way to transform to be a well-posed.
1.1 What are ill-posed problems?
According to the definition given by Jacques Hadamard (1865-1963),
the mathematical term well-posed problem has criteria as:
• a solution exists,
• the solution is unique,
• the solution’s behavior changes continuously with the initial
conditions.
4
R. Carter "Exploring Consciousness" (2002)
A psychological test for children's intelligence Fighting ?
Playing ?
Pushing game?
… Not unique!
(multiple answers
are possible)
5. 1.2 A standard / benchmark to test intelligence
5
“Where is a dog ?”
Mental rotation
Story telling
Douglas R. Hofstadter (1999)
introduced Bongard Problems
(BPs), a set of 100 puzzles by a
M.M.Bongard (mid-1960)
Puzzles for children (or IQ tests)
"Gödel, Escher,
Bach" (1999)
A Benchmark Test
BP #47
6. Bongard problems (BPs) as a benchmark
6
Human Intelligence
Input
(ill-posed
question)
Output
(right answer as
unique solution)
Find a simple rule
to discriminate
between left and
right side patterns
Left-side Right-side
On the left: triangle is INSIDE circle
On the right: circle is INSIDE triangle
9. IQ test and BPs: difficulty level is unclear
9
0
20
40
60
80
0
10
20
30
1 11 21 31 41 51 61 71 81 91
#SolvedPersons
Performance in human subject to solve BPs
# Solved
persons
Solved
speed
SolvedSpeed[s]
https://saylordotorg.github.io/text_introduction-to-
psychology/s13-intelligence-and-language.html
→High intelligenceLow ←
Conventional IQ test
have a specific
tendency, but the
difficulty level of
BPs is unclear.
→High intelligence?
H. Foundalis (2006)
10. 1.3 Research objective:
10
• What kind of logical architecture does
solve BPs?
• Why some BPs are difficult to solve?
three shapes
or
Four white space
or
Three Triangles
or
a Circle on the left
or
one upward pointing triangle
or
two large shapes and two small shapes
or
two Triangles with same kind of shape
:
BP #21 Box-L6
Color
Size
Shape
Dependent axes
Independent axes
(Infinate)
(Infinate)
“The black
triangle is
only one in
the right”
Simple Logic
Infinite
combination
Theoretical (psychological) perspective
Engineering perspective
11. Machine intelligence
Levels of practical realization
DataQuantity/Complexity
Kalman/Particle filters,
Artificial potential method,…
(Trajectory generation, obstacle avoidance)
Brain-inspired
intelligence
(Fundamental)(Industrial/Engineering)
(Bigdata)(Somedata)
Deep Learning,
Machine Learning
(Bayesian,
probabilistic)
High reliability
if data is massive
Social,
Emotional,
Contextual
Intelligence
or Qualia,
Consciousness
Industrial/Mobile Robots
Ontology, Semantic Web
(Logical Reasoning, Knowledge-based)
(Natural)
On-line & Adaptive control
Data-driven AI
Knowledge-based AI
11
12. It is possible to treat them
in a common way in logic
My hypothesis
1.4 Mathematical formulation in Logic
12
30/31* 0/31*
* H. Foundalis (2006)
Left-side Right-side Left-side Right-side
Easy Difficult
(Almost) everyone solved No one solved.
#73#23
Difficult to treat them in a
common procedure in pattern
matching/classification methods
13. 1.4 Mathematical formulation in Logic
13
30/31* 0/31*
* H. Foundalis (2006)
Left-side Right-side Left-side Right-side
Easy Difficult
(Almost) everyone solved No one solved.
#73#23
𝐿𝑖 ∈ 𝑃𝐴, 𝑅𝑖 ∈ 𝑃𝐵, 𝑖 = {1,2,3,4,5,6}
(𝐿𝑖∈ 𝑃𝐴) ⋂ 𝑅𝑖 ∈ 𝑃𝐵 ⋂ (𝐿𝑖∉ 𝑃𝐵) ⋂ (𝑅𝑖∉ 𝑃𝐴) ⋂ (PA ⋂ PB= ɸ)
SA ⋂ SB= ɸ
Theory of Logic
Rule: find properties to satisfy that
14. Problem definition
14
Independent Properties
Circle1: { Size, Texture, Color,
Shape Characteristics, position with
respect to the box...};
Triangle1:..
Dependent Properties
Circle1: { Smaller than Triangle1, To
the bottom of...};
Triangle1:..
Properties (xL1)
= ({independent properties}objects ,{dependent properties}within objects )
15. SA ⋂ SB= ɸ
n: number of features
Objects on left and right side
xL1= (x1
L1, x2
L1….. xn
L1 ……)
xR1= (x1
R1, x2
R1….. xn
R1……)
Properties (xL1) = ({in. properties}objects ,{dep. properties}within objects )
BP as an isolated area in a subspace
of the space spanned by infinite
properties
P= {P1, P2….. Pn……}
xL1= (Circle, Square, …)
Circle = (Size, Texture,… Smaller than..,)
Triangle = (Size, Texture,… Bigger than..,)
Side A
(L1)
15
Left-side Right-side
L1
L2
L3
L4
L5
L6
R1
R2
R3
R4
R5
R6
Side A Side B
16. xL1= (Circle, Square, …)
Side A
(L1)
Condition that satisfies BP solution
Required rule to check for dissimilarity (solution)
16
Properties (xL1) = ({in. properties}objects ,{dep. properties}within objects )
xL1= (Circle, Square, …)
Circle = (Size, Texture,… Smaller than..,)
Triangle = (Size, Texture,… Bigger than..,)
Side A
(L1)
SA ⋂ SB= ɸ
(𝐿𝑖∈ 𝑃𝐴) ⋂ 𝑅𝑖 ∈ 𝑃𝐵 ⋂ (𝐿𝑖∉ 𝑃𝐵) ⋂ (𝑅𝑖∉ 𝑃𝐴) ⋂ (PA ⋂ PB= ɸ)
||(𝐿𝑖 ∈ 𝑃𝐴)^ 𝑅𝑖 ∈ 𝑃𝐵 ^ (𝐿𝑖∉ 𝑃𝐵)^(𝑅𝑖∉ 𝑃𝐴) ||
→ ||(𝐿𝑖, ℎ𝑎𝑠, 𝑃𝐴) ^ (𝑅𝑖, ℎ𝑎𝑠, 𝑃𝐵)||
equivalent
17. M.M. Bongard
1970
2006
1975
1993
D. R. Hofstadter
Maksimov
problems
(MP’s)
V.V. Maksimov
Adaptive
concept
learning
algorithm
(RF4)
H. Foundalis
Phaeaco
1967 17
1.5 Past work solved limited BPs
Concept
networks
Puzzle
design
18. M.M. Bongard
1970
1975
1995
D. R. Hofstadter
Maksimov
problems
(MP’s)
V.V. Maksimov
Adaptive
concept
learning
algorithm
(RF4)
1967 18
1.5 Past work solved limited BPs
Concept
networks
Puzzle
design
2006
H. Foundalis
• Concept networks
• Frames
• Meta data
• Filters
• Sameness detector
Opposite
Right
Left
Up
Down
Similar
High
Low
4
Composed
of
Line
SegmentSquare
3
19. M.M. Bongard
1970
2006
1975
1995
D. R. Hofstadter
Maksimov
problems
(MP’s)
V.V. Maksimov
Adaptive
concept
learning
algorithm
(RF4)
Harry Foundalis
Phaeaco
1967 19
1.5 Past work solved limited BPs
Concept
networks
Puzzle
design
MP
20. M.M. Bongard
1970
2006
1975
1995
D. R. Hofstadter
V.V. Maksimov
Adaptive
concept
learning
algorithm
(RF4)
Harry Foundalis
Phaeaco
1967 20
1.5 Past work solved limited BPs
Concept
networks
Puzzle
design
⩝ B ∈ boxes, ⩝ X ∈
Btexture(X) = white →
class 1
21. M.M. Bongard
1970
2006
1975
1995
D. R. Hofstadter
Maksimov
problems
(MP’s)
V.V. Maksimov
Adaptive
concept
learning
algorithm
(RF4)
Harry Foundalis
Phaeaco
1967 21
1.5 Past work solved limited BPs
Concept
networks
Puzzle
design
22. 2. Semantic Web Technology
22
What is ontology ?
Leaf Green
is
<Subject , Predicate , Object>
Jisha is a girl who lives in Japan.
Jisha Girl
is
Human
LivesIn
Japan
is
W3C standard
Ontology design is important
23. 3. A RDF-based knowledge
representation to solve a typical BP
23
• Concept networks
• Frames
• Meta data
• Filters
• Sameness detector
D. R. Hofstadter’s idea:
Theory of Logic based on Ontology
My proposal
• Knowledge Base Representation
• ABox+Tbox Structure
• Template Design
• SPARQL Query Form
• SWRL Rules
24. Proposed Architecture
24
To treat multi-level of abstraction using the concepts like- concept
network, meta-descriptor, sameness detector, filtering, and frames.
ABoxConcept
network
Sameness
Detector
And Filtering
Pruning of
Knowledge
TBox
33. SWRL rules
33
Rule for Cross checking dissimilarity :
Rule 𝑅′
satisfies (𝐿1, 𝐿2, 𝐿3, 𝐿4, 𝐿5, 𝐿6)
Rule 𝑅′
consistent in Group A
(Inference -> Group A PredicateA Object1)
Rule R satisfies (𝑅1, 𝑅2, 𝑅3, 𝑅4, 𝑅5, 𝑅6)
Rule R consistent in Group B
(Inference-> Group B PredicateA Object2)
(𝐿1 PredicateA Object1)(𝑅1 PredicateA Object2)
If (Object1 isSameAs Object2)
Rule 𝑅′ and Rule R are not consistent for 𝐿1 and 𝑅1 sides respectively
If (Object1 DifferentFrom Object2)
Rule 𝑅′
and Rule R is consistent for Left and Right sides respectively
(Left Has Infered Shape ?a) ^
(Right Has Infered Shape ?ab)
(Left Consists of Shape ?a) ^
(Right Consists of Shape ?ab) ^
(?a Is Different From ?ab)
→
34. SWRL Rules and Querying
55 SWRL rules were used to generate minimum 12 new RDF inferred data.
(first level of inference (minimum 12) ->
second level of inference (minimum 2) (maximum 4)
Among these 55 SWRL rules-
32 rules were used as first level of inference (for similarity check)
23 rules were used as second level of inference (for dissimilarity check)
Example-
⩝x ⋿y1 ⋿y2 ⋿y3 ⋿y4 ⋿y5 ⋿y6 ⋿x1 ⋿x2 ⋿x3 ⋿x4 ⋿x5 ⋿x6
(consists_of_shape(x,polygon) ⟺
x(x1)˄x(x2) ˄x(x3) ˄x(x4) ˄x(x5) ˄x(x6) ˄(has(x1,y1)) ˄(has(x2,y2))
˄(has(x3,y3)) ˄(has(x4,y4)) ˄(has(x5,y5)) ˄(has(x6,y6))
˄(isa(y1,setoflines)) ˄ (isa(y2,setoflines)) ˄ (isa(y3,setoflines)) ˄
(isa(y4,setoflines)) ˄ (isa(y5,setoflines)) ˄ (isa(y6,setoflines))
˄(hastexture(x1,closed_shaped)) ˄(hastexture(x2,closed_shaped))
˄(hastexture(x3,closed_shaped)) ˄(hastexture(x4,closed_shaped))
˄(hastexture(x5,closed_shaped)) ˄(hastexture(x6,closed_shaped)) 34
40. Proposed description logic architecture to solve 65 BPs out of 100 BPs
40
For reasoning a set of 55 SWRL rules
were used to generate 12 new RDF
inferred data. Among these 55 SWRL
rules, 32 rules were used as first level
of inference (for similarity check) and
the rest 23 rules were used as second
level of inference (for dissimilarity
check) to find solution to a give BP.
Our proposed framework could solve
65 BPs out of the 100 original BPs.
The inferred knowledge of each BP
undergoes three-level of regressive
funneling and pruning approach (i.e.-
SPARQL query, SWRL based first
level of inference and SWRL based
second level of inference). Each stage
notices a reduction in the predicted
outcome of the selected BP.
41. Result/Comparison with past work-
41
Result of our proposed model
Here the border lines are given as a tentative borders to determine “Moderate BP”
and “Difficult BP”
Comparison with other work
43. BP#38
43
circle < triangle
(size)
triangle < circle
(size)
Right answer: Right answer:
5. Solving BPs with Dependent Properties
Towards understanding the relational properties
among each objects in a given BP
Properties={IP,DP}
44. 5. Solving BPs with Dependent Properties
44
Class(set)
Instance (element)
Concept
Actual object
Rules can apply to Concept, but not to each element.
45. 45
BPs using Independent properties BPs using dependent properties
Matches with Piaget’s Theory of Cognitive Development
[four cognitive development stage in children]
48. 48
Inferred solution for BP#38 using Protege
circle < triangle
(size)
triangle < circle
(size)
Right answer: Right answer:
49. Inferred solution for BP#38 using Protege
49
circle < triangle
(size)
triangle < circle
(size)
Right answer: Right answer:
50. Discussion and Conclusion (1/2)
• Our proposed framework could solve 65 BPs out of the 100 BPs. The inferred
knowledge of each BP undergoes three-level of regressive funneling and pruning
(SPARQL query, SWRL based first level of inference and SWRL based second
level of inference).
• We have proved that our model with RDF based knowledge base is efficient in
solving BPs (Ill-posed problems). By considering a very simple logic-
• Hence an ontology based approach for solving ill-posed problems could be a step
towards brain inspired general AI.
• We have validated the hypothesis of Dr. D. R. Hofstadter’s of using Concept
networks, frames, meta data and sameness detector as a possible step towards
solving BPs.
50
(𝐿𝑖∈ 𝑃𝐴) ⋂ 𝑅𝑖 ∈ 𝑃𝐵 ⋂ (𝐿𝑖∉ 𝑃𝐵) ⋂ (𝑅𝑖∉ 𝑃𝐴) ⋂ (PA ⋂ PB= ɸ)
||(𝐿𝑖 ∈ 𝑃𝐴)^ 𝑅𝑖 ∈ 𝑃𝐵 ^ (𝐿𝑖∉ 𝑃𝐵)^(𝑅𝑖∉ 𝑃𝐴) ||
→ ||(𝐿𝑖, ℎ𝑎𝑠, 𝑃𝐴) ^ (𝑅𝑖, ℎ𝑎𝑠, 𝑃𝐵)||
51. Discussion and Conclusion (2/2)
• In the future work, this framework can be embedded in the hybrid system as an
automatic BP solver changing analogies associated with vision-based analyzers for
spatial representation. It can open the new horizon of the logical reasoning system
to incorporate data-driven models for decision making process in the dynamic
environment.
• As a real-world application, this approach can help us in understanding the
difference between cancerous cells (DNC) (with cell description-irregular shape,
protoplasm shape-circle, stripped texture..and so forth) and non-cancerous cells
(DNN) (with cell description-circular shape, protoplasm shape-square, shaded
texture..and so forth).
51
Data-driven Methods,
Clustering of Data
Data Model, Data Structure
(Model-based Approach)(Model-Free Approach)
52. Application of Ontology-
1. Ontology Scheme in Agriculture Application
2. Human assistance for driving automation
52
53. 1. Ontology Scheme in Agriculture Application
Need for plant automation-
• Decline in labour force (aging society)
• Regular monitoring (being alert for sudden changes
in conditions) with expert human knowledge
53
Collection of data
[sensors-IR…etc]
Collection of expert
knowledge (on plant
growth and related
condition)
Tomato
Ontology with
SWRL rules
SWRL rules
based Inference
Inference
Automation unit
(to add the
deficient fertilizer)
Converting
Sensor data to
CSV data
CSV
sensor data
to RDF
data
+
Nitrogen
Approach-
55. References
[1] A. Linhares, A glimpse at the metaphysics of Bongard problems, Artificial Intelligence, vol. 121, no.
1-2, pp. 251-270, 2000.
[2] S. Kazumi, N. Ryohei, Adaptive concept learning algorithm: RF 4, Transactions of Information
Processing Society, Vol. 36, No. 4, pp. 832 - 839, 1995.
[3] J. Hernández-Orallo, F. Martínez-Plumed, U. Schmid, M. Siebers and D. Dowe, Computer models
solving intelligence test problems: Progress and implications, Artificial Intelligence, vol. 230, pp. 74-107,
2016.
[4] H. Foundalis, (2006). Phaeaco: A Cognitive Architecture Inspired by Bongard’s Problems. Doctoral
dissertation, Indiana University, Center for Research on Concepts and Cognition (CRCC), Bloomington,
Indiana.
[5] D. R Hofstadter, (1979). Gödel, Escher, Bach: an Eternal Golden Braid. New York: Basic Books.
[6] S. Durbha and R. King, Semantics-enabled framework for knowledge discovery from Earth
observation data archives, IEEE Transactions on Geoscience and Remote Sensing, vol. 43, no. 11, pp.
2563-2572, 2005.
[7] A. Maarala, X. Su and J. Riekki, Semantic Reasoning for Context-Aware Internet of Things
Applications, IEEE Internet of Things Journal, vol. 4, no. 2, pp. 461-473, 2017.
[8] M. Zand, S. Doraisamy, A. Abdul Halin and M. Mustaffa, Ontology-Based Semantic Image
Segmentation Using Mixture Models and Multiple CRFs, IEEE Transactions on Image Processing, vol.
25, no. 7, pp. 3233-3248, 2016.
[9] K. Salameh, J. Tekli and R. Chbeir, SVG-to-RDF Image Semantization, Similarity Search and
Applications, pp. 214-228, 2014. 55
56. Publications
[CONFERENCE]
• Jisha Maniamma and Hiroaki Wagatsuma (2018): How We Treat Logical Rules to Solve Puzzles: A Semantic Web
Approach for Bongard Problems, 日本神経回路学会 第28回全国大会(JNNS2018)Posters & Demos, The 28th
Annual Conference of the Japanese Neural Network Society (JNNS 2018), October 24 - 27, 2018, Okinawa Institute
of Science and Technology (OIST), Okinawa, Japan.
• Jisha Maniamma and Hiroaki Wagatsuma (2018): A Semantic Web Technique as Logical Inference Puzzle-Solver
for Bongard Problems, ISWC 2018 Posters & Demos, The 17th International Semantic Web Conference (ISWC
2018), October 8 - 12, 2018 Monterey, California, USA.
• Jisha Maniamma and Hiroaki Wagatsuma (2018): Human Abduction for Solving Puzzles to Find Logically
Explicable Rules to Discriminate Two Picture Groups Ostracized Each Other: An Ontology-based Model, FAIM
Workshop On Architectures And Evaluation For Generality, Autonomy & Progress in AI, 27th International Joint
Conference on Artificial Intelligence and the 23rd European Conference on Artificial Intelligence (IJCAI-ECAI
2018), July 15, 2018, Stockholm, Sweden.
• Jisha Maniamma and Hiroaki Wagatsuma (2017): An Ontology-Based Knowledge Representation Towards Solving
Bongard Problems, The 12th International Conference on Innovative Computing, Information and Control (ICICIC
2017), August 28–30, 2017, Kurume, Japan.
• Maniamma, J., Hagio, M., Togo, M., Shimotake, A., Matsumoto, R., Ikeda, A., Wagatsuma, H. (2017): A High-
Precision Skilled Movement Evaluation by using Curvature Analysis in the Simultaneous Recording of 3D Motion
Capture System and Intracranial Video-EEG Monitoring and Stimulation, The 39th Annual International Conference
of the IEEE Engineering in Medicine and Biology Society (EMBC 2017), ID FrDT17-08.3, July 14, 2017, JEJU
International Convention Centre, Jeju Island, Korea.
• Jisha Maniamma and Hiroaki Wagatsuma (2017): Semantic-Web Based Representations to Solve Bongard
Problems with a Logical Reasoning Architecture, 日本神経回路学会全国大会講演論文集(JNNS 2017), 27:
71‐72, Sep. 20, 2017.
56
57. [Journal PUBLICATIONS]
Jisha Maniamma and Hiroaki Wagatsuma, An Semantic Web-based Representation of
Human-logical Inference for Solving Bongard Problems, Journal of Universal Computer
Science: Special Issue on “New Trends in Logic Reasoning Based Decision Making.”, in
press, 2020.
Jisha Maniamma and Hiroaki Wagatsuma, An ontology-based knowledge representation
towards solving Bongard problems, ICIC Express Letters: an International Journal of
Research and Surveys, 12(7): 681-688, 2019.
57
Jisha Maniamma and Hiroaki Wagatsuma, A Semantic Web Technique as Logical
Inference Puzzle-Solver for Bongard Problems, Proceedings of the ISWC 2018 Posters &
Demonstrations, Industry and Blue Sky Ideas Tracks co-located with 17th International
Semantic Web Conference (ISWC 2018), Monterey, USA, October 8th to 12th, 2018.
Editor's Notes
here I will take you through the journey is taken by many researchers towards solving this issue
Bongard problems (Bongard analogies) were formulated in the year 1967 by MM Bongard ( a Russian scientist) in his book the titled ”problem of recognition”. They are are a set of 100 puzzles. MM Bongard was intersected in the automation of visual perception. But it was little known about these puzzles to the western world till 1970.
here I will take you through the journey is taken by many researchers towards solving this issue
Bongard problems (Bongard analogies) were formulated in the year 1967 by MM Bongard ( a Russian scientist) in his book the titled ”problem of recognition”. They are are a set of 100 puzzles. MM Bongard was intersected in the automation of visual perception. But it was little known about these puzzles to the western world till 1970.
here I will take you through the journey is taken by many researchers towards solving this issue
Bongard problems (Bongard analogies) were formulated in the year 1967 by MM Bongard ( a Russian scientist) in his book the titled ”problem of recognition”. They are are a set of 100 puzzles. MM Bongard was intersected in the automation of visual perception. But it was little known about these puzzles to the western world till 1970.
here I will take you through the journey is taken by many researchers towards solving this issue
Bongard problems (Bongard analogies) were formulated in the year 1967 by MM Bongard ( a Russian scientist) in his book the titled ”problem of recognition”. They are are a set of 100 puzzles. MM Bongard was intersected in the automation of visual perception. But it was little known about these puzzles to the western world till 1970.
here I will take you through the journey is taken by many researchers towards solving this issue
Bongard problems (Bongard analogies) were formulated in the year 1967 by MM Bongard ( a Russian scientist) in his book the titled ”problem of recognition”. They are are a set of 100 puzzles. MM Bongard was intersected in the automation of visual perception. But it was little known about these puzzles to the western world till 1970.