Get documents about "1-introduction.pptx
Document Sample
1/25/2010
Intelligent Systems
COMSM0301: 2009/0 - Lecture 1
Ghost in the Shell
“Introduction to Learning from Synthetic
Structured Data”
Deep Blue
Oliver Ray REASONING
HAL 9000
(oray@cs.bris.ac.uk)
Department of Computer Science
University of Bristol
Analytic
25th January, 2010 Robot Scientist
functional REPRESENTATION relational
State-of-the-Art A Hierarchy of Representations
Synthetic
ML
REASONING
DM ATTRIBUTE- MULTIPLE MULTI INDUCTIVE
VALUE INSTANCE RELATIONAL LOGIC
PA LEARNING LEARNING LEARNING PROGRAMMING
Analytic O KR (AVL) (MIL) (MRL) (ILP)
functional REPRESENTATION relational functional REPRESENTATION relational
Database Perspective A Hierarchy of Reasoning
Vector Spreadsheet relational deductive Generalisation:
database database Synthetic
INDUCTION From particular cases
To general laws
Explanation:
REASONING ABDUCTION From observed effects
To hidden causes
ATTRIBUTE- MULTIPLE MULTI INDUCTIVE
VALUE INSTANCE RELATIONAL LOGIC Consequence:
LEARNING LEARNING LEARNING PROGRAMMING DEDUCTION From given knowledge
Analytic
(AVL) (MIL) (MRL) (ILP)
To necessary implications
functional REPRESENTATION relational
1
1/25/2010
Syllogistic Perspective Learning From Structured Data
These beans are white
Synthetic Synthetic LSD
INDUCTION These beans are from this bag
All beans from this bag are white
ILP
All beans from this bag are white ML
REASONING ABDUCTION These beans are white REASONING
These beans are from this bag
ALP
DM
All beans from this bag are white LP
DEDUCTION These beans are from this bag PA
Analytic Analytic O
These beans are white
KR
functional REPRESENTATION relational
LSD: What? LSD: Why?
Extend the representation power of conventional machine
learning systems to support real-world relational data (with
Supervised Machine Learning internal structure and external relationships)
Extend the reasoning power of conventional knowledge
for representation frameworks to support real-world synthetic
inference (for uncertain and incomplete knowledge)
Structured Knowledge Representations
Focus on the key challenges in representation and reasoning
overlooked by conventional approaches which assume useful
features and relevant knowledge is known in advance
LSD: How? The Start: Learning as Func. Approx.
1 upgrade learner
Synthetic LSD
2 downgrade representation
I O
e t
3
REASONING hybridise
inference Input Space Output Space
H
(examples) (targets)
m
Analytic Hypothesis Space
(models)
functional REPRESENTATION relational
2
1/25/2010
The End: Learning with Uncertain Info. Course Administration
Probabilistic Logic Learning (PLL) Lecturer
Oliver Ray MVB 2.xx oray@cs.bris.ac.uk
Attendance
Week 13 QB 1.69 Mon-Fri 9am - 1pm (4h)
Week 14 QB 1.69 Mon-Thur 2pm - 5pm (3h)
(i) entailment (ii) interpretations (iii) proofs Assessment
Assignment 1 20 % Week 14
Nonmonotonic Inductive Logic Programming (nmILP) Assignment 2 30 % Week 21
Examination 50 % May/June
Foundations
Intro. to Machine Learn. COMS30301
Artif. Intell. & Logic Prog. COMS30106
(i) inverse entailment (ii) least generalisation (iii) inverse resolution
Course Timetable Course Objectives
Mon Tue Wed Thu Fri
25/01 26/01 27/01 28/01 29/01
09:00 – 09:50 lec lec lec lec lec Appreciate the key limitations of traditional (attribute-value)
10:00 – 10:45 tut tut tut tut tut Machine Learning methods and understand the practical
need to overcome these limitations
11:00 – 11:50 lec lec lec PROJ lec
12:00 – 12:30 cwk cwk cwk PROJ cwk
Mon Tue Wed Thu Fri
Appreciate the key methods for learning with expressive
01/02 02/02 03/02 04/02 05/02 (relational, logical & probabilistic) representations and
understand the trade-offs such methods bring about
14:00 – 14:50 lec lec LAB lec CNS
15:00 – 15:45 tut tut LAB tut CNS
16:00 – 16:30 cwk cwk cwk cwk CNS
Course Reading Case Study: Mutagenisis
Relational Inductive Logic Programming: Mutagenic compounds encourage the mutation of DNA and pose
Data Mining Techniques and Applications serious health risks which play a key role in drug development
The mutagenic activity of many compounds is known from in-vitro
studies, but it is not practical to test all compounds in this way
Thus we want predictive models, or Structure Activity Relationships
(SARs), that relate mutagenicity to physiochemical properties
The mutagenesis data set contains 230 Aromatic and Heteroaromatic
Nitro compounds, each described by 1 class label and 4 attributes:
act - mutagenic activity (log TA98 Ames test)
εLUMO - energy of lowest unoccupied molecular orbital
logP - hydrophobicity (log octanol/water coecient)
Ia - indicator for an acenthrylene
I1 - indicator for 3 or more fused rings
Foundations of Inductive Logical and Relational Learning The data set usually split into 188 “regression friendly” compounds
Logic Programming
and 42 “regression unfriendly” compounds
3
1/25/2010
Mutagenisis as Attribute-Value Learning Visualisation
BiLinear Regression Mutagenisis as Attribute-Value Learning
Basic Chemical Structures A New Structural Alert
4
1/25/2010
The Bottom Line Tutorial 1: Represent This!
Most real-world data is highly structured: sets, lists, trees, graphs, Molecule Chemical Structure Class
space, time, and complex relationships within and between objects
H
formaldehyde C O neg
H
To use attribute-value learners, the structure of the data must be
collapsed by extracting a predetermined set of features H
methane H C H pos
As there are a potentially infinite number of features, the data itself is H
often the most compact and complete representation H H
C C
benzene H C C H neg
In reality, identifying relevant features is the key learning problem; C C
and new techniques are needed to learn with structured data H H
Molecules Schema Molecules Database
Molecule: Atom: Bond:
Name Molecule Class Name Class Atom-ID Element Bond-ID Valency
formaldehyde neg atm_form_1 C bnd_form_1 2
1 methane pos atm_form_2 O bnd_form_2 1
benzene neg atm_form_3 H bnd_form_3 1
Contains
atm_form_4 H bnd_meth_1 1
atm_meth_1 C … …
* … … bnd_benz_12 1
Atom-ID Atom Element Contains: atm_benz_12 H
1 1 Molecule Atom Connects:
formaldehyde atm_form_1 Bond Atom1 Atom2
Connects formaldehyde atm_form_2 bnd_form_1 atm_form_1 atm_form_2
formaldehyde atm_form_3 bnd_form_2 atm_form_1 atm_form_3
? formaldehyde atm_form_4 bnd_form_3 atm_form_1 atm_form_4
methane atm_meth_1 bnd_meth_1 atm_meth_1 atm_meth_2
Bond-ID Bond Valency … … … … …
benzene atm_benz_12 bnd_benz_12 atm_benz_11 atm_benz_12
Simplified Molecules Database Tutorial 2: Represent This!
Molecule: Atom:
Name Class Atom Molecule Element
formaldehyde neg atm_form_1 formaldehyde C
1. TRAINS GOING EAST 2. TRAINS GOING WEST
methane pos atm_form_2 formaldehyde O
benzene neg atm_form_3 formaldehyde H 1. 1.
atm_form_4 formaldehyde H
atm_meth_1 methane C 2. 2.
… … …
Bond: atm_benz_12 benzene H
3. 3.
Atom1 Atom2 Valency
4. 4.
atm_form_1 atm_form_2 2
atm_form_1 atm_form_3 1 5. 5.
atm_form_1 atm_form_4 1 (i) Bond becomes a weak entity
atm_meth_1 atm_meth_2 1 (ii) Contains and Connects are both
… … … treated as attributes
atm_benz_11 atm_benz_12 1
5
