the road to intelligence is paved with a million million expert systems christine alvarado 6...
DESCRIPTION
OK, So Why Is this Dangerous? What is intelligence? Many ideas: passing the Turing Test, displaying common sense, etc. Perception What is knowledge? Specific facts about the world e.g. Humans have two legs, breakfast is served in the morning, etc.TRANSCRIPT
![Page 1: The Road to Intelligence Is Paved with a Million Million Expert Systems Christine Alvarado 6 November 2002](https://reader035.vdocuments.site/reader035/viewer/2022062401/5a4d1b467f8b9ab0599a350a/html5/thumbnails/1.jpg)
The Road to Intelligence Is Paved with a Million Million Expert Systems
Christine Alvarado6 November 2002
![Page 2: The Road to Intelligence Is Paved with a Million Million Expert Systems Christine Alvarado 6 November 2002](https://reader035.vdocuments.site/reader035/viewer/2022062401/5a4d1b467f8b9ab0599a350a/html5/thumbnails/2.jpg)
A Dangerous Idea?
You need knowledge to be intelligent
![Page 3: The Road to Intelligence Is Paved with a Million Million Expert Systems Christine Alvarado 6 November 2002](https://reader035.vdocuments.site/reader035/viewer/2022062401/5a4d1b467f8b9ab0599a350a/html5/thumbnails/3.jpg)
OK, So Why Is this Dangerous?
What is intelligence?Many ideas: passing the Turing Test,
displaying common sense, etc.Perception
What is knowledge?Specific facts about the worlde.g. Humans have two legs, breakfast is
served in the morning, etc.
![Page 4: The Road to Intelligence Is Paved with a Million Million Expert Systems Christine Alvarado 6 November 2002](https://reader035.vdocuments.site/reader035/viewer/2022062401/5a4d1b467f8b9ab0599a350a/html5/thumbnails/4.jpg)
A More Dangerous Idea
We should explicitly represent specific knowledge within perceptual computer systems
Questions I will address:What does this mean?Why should we do this?How can we do this?
![Page 5: The Road to Intelligence Is Paved with a Million Million Expert Systems Christine Alvarado 6 November 2002](https://reader035.vdocuments.site/reader035/viewer/2022062401/5a4d1b467f8b9ab0599a350a/html5/thumbnails/5.jpg)
Data-Driven vs. Concept-Driven Perception
Conceptual Information
Sensory Input
![Page 6: The Road to Intelligence Is Paved with a Million Million Expert Systems Christine Alvarado 6 November 2002](https://reader035.vdocuments.site/reader035/viewer/2022062401/5a4d1b467f8b9ab0599a350a/html5/thumbnails/6.jpg)
The Importance of Context
![Page 7: The Road to Intelligence Is Paved with a Million Million Expert Systems Christine Alvarado 6 November 2002](https://reader035.vdocuments.site/reader035/viewer/2022062401/5a4d1b467f8b9ab0599a350a/html5/thumbnails/7.jpg)
The Importance of Context
![Page 8: The Road to Intelligence Is Paved with a Million Million Expert Systems Christine Alvarado 6 November 2002](https://reader035.vdocuments.site/reader035/viewer/2022062401/5a4d1b467f8b9ab0599a350a/html5/thumbnails/8.jpg)
The Importance of Context
![Page 9: The Road to Intelligence Is Paved with a Million Million Expert Systems Christine Alvarado 6 November 2002](https://reader035.vdocuments.site/reader035/viewer/2022062401/5a4d1b467f8b9ab0599a350a/html5/thumbnails/9.jpg)
The Importance of Context
![Page 10: The Road to Intelligence Is Paved with a Million Million Expert Systems Christine Alvarado 6 November 2002](https://reader035.vdocuments.site/reader035/viewer/2022062401/5a4d1b467f8b9ab0599a350a/html5/thumbnails/10.jpg)
Another Example
![Page 11: The Road to Intelligence Is Paved with a Million Million Expert Systems Christine Alvarado 6 November 2002](https://reader035.vdocuments.site/reader035/viewer/2022062401/5a4d1b467f8b9ab0599a350a/html5/thumbnails/11.jpg)
Another Example
2) 1 + 3 = 4
1) 1 + 2 = 2
3) 5 + 6 = 10
![Page 12: The Road to Intelligence Is Paved with a Million Million Expert Systems Christine Alvarado 6 November 2002](https://reader035.vdocuments.site/reader035/viewer/2022062401/5a4d1b467f8b9ab0599a350a/html5/thumbnails/12.jpg)
Another Example
![Page 13: The Road to Intelligence Is Paved with a Million Million Expert Systems Christine Alvarado 6 November 2002](https://reader035.vdocuments.site/reader035/viewer/2022062401/5a4d1b467f8b9ab0599a350a/html5/thumbnails/13.jpg)
Where’s the Knowledge?
Conceptual KnowledgeCows, sheep and pigs are
all barnyard animalsStructural Knowledge
Cows’ eyes are above their noses
Cows have spotsCows have long noses
Not included in most recognizers
Only implicitly included in most recognizers
We should represent this knowledge explicitly!
![Page 14: The Road to Intelligence Is Paved with a Million Million Expert Systems Christine Alvarado 6 November 2002](https://reader035.vdocuments.site/reader035/viewer/2022062401/5a4d1b467f8b9ab0599a350a/html5/thumbnails/14.jpg)
Feature-Based Models
Feature Extraction
Structural Knowledge hidden within network Which part of the
network represents “cows have legs”
Conceptual Knowledge absent
![Page 15: The Road to Intelligence Is Paved with a Million Million Expert Systems Christine Alvarado 6 November 2002](https://reader035.vdocuments.site/reader035/viewer/2022062401/5a4d1b467f8b9ab0599a350a/html5/thumbnails/15.jpg)
![Page 16: The Road to Intelligence Is Paved with a Million Million Expert Systems Christine Alvarado 6 November 2002](https://reader035.vdocuments.site/reader035/viewer/2022062401/5a4d1b467f8b9ab0599a350a/html5/thumbnails/16.jpg)
Why Make Knowledge Explicit?
To include conceptual knowledge in recognition
To understand why the computer system is making a mistake
To allow humans to construct recognizers
![Page 17: The Road to Intelligence Is Paved with a Million Million Expert Systems Christine Alvarado 6 November 2002](https://reader035.vdocuments.site/reader035/viewer/2022062401/5a4d1b467f8b9ab0599a350a/html5/thumbnails/17.jpg)
Including Conceptual Knowledge
Feature-based approach cannot easily handle conceptual knowledgeA picture of a cow encodes
structural knowledgeBut how do you
incorporate “cows and sheep are barnyard animals”?
![Page 18: The Road to Intelligence Is Paved with a Million Million Expert Systems Christine Alvarado 6 November 2002](https://reader035.vdocuments.site/reader035/viewer/2022062401/5a4d1b467f8b9ab0599a350a/html5/thumbnails/18.jpg)
Contextual Knowledge is Essential
Contextual knowledge determines the interpretation for this shape:
![Page 19: The Road to Intelligence Is Paved with a Million Million Expert Systems Christine Alvarado 6 November 2002](https://reader035.vdocuments.site/reader035/viewer/2022062401/5a4d1b467f8b9ab0599a350a/html5/thumbnails/19.jpg)
Once We Have Concepts, We Can Generalize
Face
eyes
nose
mouth
Eyes
Nose
Mouth
![Page 20: The Road to Intelligence Is Paved with a Million Million Expert Systems Christine Alvarado 6 November 2002](https://reader035.vdocuments.site/reader035/viewer/2022062401/5a4d1b467f8b9ab0599a350a/html5/thumbnails/20.jpg)
More Human Understanding
The system can explain the reasons for its beliefs
I think the shape is an arrow even though it only sort of
looks like one because this is a finite state machine.
![Page 21: The Road to Intelligence Is Paved with a Million Million Expert Systems Christine Alvarado 6 November 2002](https://reader035.vdocuments.site/reader035/viewer/2022062401/5a4d1b467f8b9ab0599a350a/html5/thumbnails/21.jpg)
Should We Get Rid of Feature-Based Recognizers?
Of course not!All these concepts have to bottom out
somewhereFeature-based recognizers very useful for
low level recognitionWe should integrate information from
feature-based recognition with explicitly represented knowledge
![Page 22: The Road to Intelligence Is Paved with a Million Million Expert Systems Christine Alvarado 6 November 2002](https://reader035.vdocuments.site/reader035/viewer/2022062401/5a4d1b467f8b9ab0599a350a/html5/thumbnails/22.jpg)
OK, This Sounds Good, But How?
Dynamically Constructed Object Oriented Bayesian NetworksCreate a fragment of a Bayesian Network
for each specific piece of knowledgeDynamically link them together as input
arrives
![Page 23: The Road to Intelligence Is Paved with a Million Million Expert Systems Christine Alvarado 6 November 2002](https://reader035.vdocuments.site/reader035/viewer/2022062401/5a4d1b467f8b9ab0599a350a/html5/thumbnails/23.jpg)
Representing Knowledge
(Define Shape And-Gate (components (Line L1) (Line L2) (Line L3) (Semi-Circle S)) (constraints (parallel L1 L2) (same-horiz-pos L1 L2) (same-length L1 L2) (connected S.p1 L3.p1) (connected S.p2 L3.p2) (meets L1.p2 L3) (meets L2.p2 L3)))
L1:L2:L3:S:
C1:C2:C3:C4:C5:C6:C7:
And-Gate
L1 L2 L3 S C1 C2 C7…
Description Network Fragment
![Page 24: The Road to Intelligence Is Paved with a Million Million Expert Systems Christine Alvarado 6 November 2002](https://reader035.vdocuments.site/reader035/viewer/2022062401/5a4d1b467f8b9ab0599a350a/html5/thumbnails/24.jpg)
Hierarchical Representation
To encode: “Often, an inverter precedes an and-gate”
(Define Shape-Composition Inv-Before-And
(components (And-Gate ag) (Inverter inv)) (constraints (connects inv ag) (precedes inv ag)))
AG:INV:
Inv-Before-And
AG INV C1 C2
And-Gate FragmentC1:C2:
![Page 25: The Road to Intelligence Is Paved with a Million Million Expert Systems Christine Alvarado 6 November 2002](https://reader035.vdocuments.site/reader035/viewer/2022062401/5a4d1b467f8b9ab0599a350a/html5/thumbnails/25.jpg)
Knowledge Acquisition
Start small: specific domainsSketches are a good place to begin
Simpler than vision2D with a temporal component
We can put some of the knowledge in ourselves
![Page 26: The Road to Intelligence Is Paved with a Million Million Expert Systems Christine Alvarado 6 November 2002](https://reader035.vdocuments.site/reader035/viewer/2022062401/5a4d1b467f8b9ab0599a350a/html5/thumbnails/26.jpg)
Conclusion
Explicit representation of knowledge……both structural and conceptual
knowledge……is a powerful way to build intelligent
systems.