Scribd
Upload
129 views19 pages

Combining Inductive and Analytical Learning: CS 5751 Machine Lear Ning Chapter 12 Comb. Inductive/Analyti Cal 1

The document discusses combining inductive and analytical learning by leveraging prior domain knowledge. It presents several approaches: KBANN initializes a neural network with the domain theory; TangentProp and EBNN incorporate prior knowledge into the objective function or hypothesis search; and FOCL alters the search operators in learning rules based on the domain theory. The goal is to develop general purpose learning that benefits from prior knowledge when available and performs well even without perfect knowledge.

Uploaded by

Arshiya Banu
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPT, PDF, TXT or read online on Scribd
129 views19 pages

Combining Inductive and Analytical Learning: CS 5751 Machine Lear Ning Chapter 12 Comb. Inductive/Analyti Cal 1

The document discusses combining inductive and analytical learning by leveraging prior domain knowledge. It presents several approaches: KBANN initializes a neural network with the domain theory; TangentProp and EBNN incorporate prior knowledge into the objective function or hypothesis search; and FOCL alters the search operators in learning rules based on the domain theory. The goal is to develop general purpose learning that benefits from prior knowledge when available and performs well even without perfect knowledge.

Uploaded by

Arshiya Banu
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPT, PDF, TXT or read online on Scribd
You are on page 1/ 19

Combining Inductive and Analytical Learning

• Why combine inductive and analytical learning?


• KBANN: prior knowledge to initialize the
hypothesis
• TangentProp, EBNN: prior knowledge alters
search objective
• FOCL: prior knowledge alters search operators

CS 5751 Machine Lear Chapter 12 Comb. Inductive/Analyti 1


ning cal
Inductive and Analytical Learning
Inductive learning Analytical learning
Hypothesis fits data Hypothesis fits domain
theory
Statistical inference Deductive inference
Requires little prior Learns from scarce data
knowledge
Syntactic inductive bias Bias is domain theory

CS 5751 Machine Lear Chapter 12 Comb. Inductive/Analyti 2


ning cal
What We Would Like
Inductive learning Analytical learning
Plentiful data Scarce data
No prior knowledge Perfect prior knowledge
• General purpose learning method:
• No domain theory  learn as well as inductive
methods
• Perfect domain theory  learn as well as PROLOG-
EBG
• Accommodate arbitrary and unknown errors in
domain theory
• Accommodate arbitrary and unknown errors in
training data
CS 5751 Machine Lear Chapter 12 Comb. Inductive/Analyti 3
ning cal
Domain Theory
Cup  Stable, Liftable, OpenVessel
Stable  BottomIsFlat
Liftable  Graspable, Light
Graspable  HasHandle
OpenVessel  HasConcavity, ConcavityPointsUp
Cup

Stable Liftable OpenVessel

Graspable

BottomIsFlat Light HasHandle HasConcavity ConcavityPointsUp


CS 5751 Machine Lear Chapter 12 Comb. Inductive/Analyti 4
ning cal
Training Examples
Cups Non-Cups
BottomIsFlat        
ConcavityPointsUp       
Expensive    
Fragile      
HandleOnTop  
HandleOnSide   
HasConcavity         
HasHandle     
Light        
MadeOfCeramic    
MadeOfPaper  
MadeOfStyroForm    
CS 5751 Machine Lear Chapter 12 Comb. Inductive/Analyti 5
ning cal
KBANN
Knowledge Based Artificial Neural Networks

KBANN (data D, domain theory B)


1. Create a feedforward network h equivalent to B
2. Use BACKPROP to tune h to fit D

CS 5751 Machine Lear Chapter 12 Comb. Inductive/Analyti 6


ning cal
Neural Net Equivalent to Domain Theory

Expensive
BottomIsFlat Stable
MadeOfCeramic
MadeOfStyrofoam
MadeOfPaper
HasHandle Graspable Liftable Cup
HandleOnTop
HandleOnSide
Light
HasConcavity OpenVessel
ConcavityPointsUp large positive weight
Fragile
large negative weight
negligible weight
CS 5751 Machine Lear Chapter 12 Comb. Inductive/Analyti 7
ning cal
Creating Network Equivalent to Domain Theory
Create one unit per horn clause rule (an AND unit)
• Connect unit inputs to corresponding clause
antecedents
• For each non-negated antecedent, corresponding
input weight w  W, where W is some constant
• For each negated antecedent, weight w  -W
• Threshold weight w0  -(n - .5) W, where n is
number of non-negated antecedents
Finally, add additional connections with near-zero
weights
Liftable  Graspable, ¬Heavy
CS 5751 Machine Lear Chapter 12 Comb. Inductive/Analyti 8
ning cal
Result of Refining the Network
Expensive
BottomIsFlat Stable
MadeOfCeramic
MadeOfStyrofoam
MadeOfPaper
HasHandle Graspable Liftable Cup
HandleOnTop
HandleOnSide
Light
HasConcavity OpenVessel
ConcavityPointsUp large positive weight
Fragile
large negative weight
negligible weight
CS 5751 Machine Lear Chapter 12 Comb. Inductive/Analyti 9
ning cal
KBANN Results
Classifying promoter regions in DNA (leave one out
testing):
• Backpropagation: error rate 8/106
• KBANN: 4/106

Similar improvements on other classification, control


tasks.

CS 5751 Machine Lear Chapter 12 Comb. Inductive/Analyti 10


ning cal
Hypothesis Space Search in KBANN
Hypothesis Space

Hypotheses that
fit training data
equally well

Initial hypothesis
for KBANN

Initial hypothesis
for Backpropagation

CS 5751 Machine Lear Chapter 12 Comb. Inductive/Analyti 11


ning cal
EBNN
Explanation Based Neural Network

Key idea:
• Previously learned approximate domain theory
• Domain theory represented by collection of neural
networks
• Learn target function as another neural network

CS 5751 Machine Lear Chapter 12 Comb. Inductive/Analyti 12


ning cal
Explanation in Terms of Domain Theory
Prior learned networks for useful concepts
combined into a single target network
Expensive
BottomIsFlat
MadeOfCeramic Stable
MadeOfStyrofoam
MadeOfPaper
HasHandle Stable
HandleOnTop
HandleOnSide Graspable Liftable Cup
Light
HasConcavity
ConcavityPointsUp
Fragile
OpenVessel

CS 5751 Machine Lear Chapter 12 Comb. Inductive/Analyti 13


ning cal
TangetProp
f(x)
Assume x, f(x) and provided as input
x xi
Modified objective for gradient descent :
  A( x ) fˆ ( x ) 
2

E   ( f ( xi )  fˆ ( xi ))  μ i  j 
2
 
j  ( x  xi )

 x x 
j
i  
 
where
A( xi )  f ( xi )
μ i  1
c

 f(x) is target function


 fˆ(x) is neural net approximation to f(x)
 A(x) is domain theory approximation to f(x)
CS 5751 Machine Lear Chapter 12 Comb. Inductive/Analyti 14
ning cal
CS 5751 Machine Lear Chapter 12 Comb. Inductive/Analyti 15
ning cal
Hypothesis Space Search in TangentProp
Hypothesis Space

Hypotheses that
Hypotheses that
maximize fit to data
maximize fit to
data and prior
knowledge

Backpropagation
TangetProp Search
Search

CS 5751 Machine Lear Chapter 12 Comb. Inductive/Analyti 16


ning cal
FOCL
• Adaptation of FOIL that uses domain theory
• When adding a literal to a rule, not only consider
adding single terms, but also think about adding
terms from domain theory
• May also prune specializations generated

CS 5751 Machine Lear Chapter 12 Comb. Inductive/Analyti 17


ning cal
Search in FOCL
Cup 

Cup  HasHandle
[2+,3-]
...
Cup  BottomIsFlat,
Cup  ¬HasHandle
Light,
[2+,3-]
Cup  Fragile HasConcavity,
[2+,4-] ConcavityPointsUp
[4+,2-]
Cup  BottomIsFlat, ...
Light, Cup  BottomIsFlat, Cup  BottomIsFlat,
HasConcavity, Light, Light,
ConcavityPointsUp, HasConcavity, HasConcavity,
HandleOnTop ConcavityPointsUp, ConcavityPointsUp,
[0+,2-] ¬ HandleOnTop HandleOnSide
[4+,0-] [2+,0-]
CS 5751 Machine Lear Chapter 12 Comb. Inductive/Analyti 18
ning cal
FOCL Results
Recognizing legal chess endgame positions:
• 30 positive, 30 negative examples
• FOIL: 86%
• FOCL: 94% (using domain theory with 76%
accuracy)
NYNEX telephone network diagnosis
• 500 training examples
• FOIL: 90%
• FOCL: 98% (using domain theory with 95%
accuracy)
CS 5751 Machine Lear Chapter 12 Comb. Inductive/Analyti 19
ning cal

You might also like