Transcript Working Group 4

Working Group 4
Creative Systems for
Knowledge Management in
Life Sciences
Purpose of this Talk

We are researching methods which we
believe could provide non-standard
solutions to complex problems

We need concrete problems to identify
possible interactions between the
working groups
Structure of Talk

Individual research directions

General techniques for creative reasoning

A case study
Computational Bioinformatics Laboratory,
Imperial College London

Progol system
– Learning of concepts in bioinformatics
– Theory behind, and implementation of ILP
– Applications:
• Predictive toxicology, secondary structure in proteins,
learning metabolic pathways

HR system
– Discovering in mathematics (and bioinformatics)
– Theory behind, and implementation of ATF
– Applications:
• Adding to databases: Integer sequences, TPTP library
• Finding invariants, inventing CSP constraints, tutorials

Scientific Discovery via integration of techniques
Centre for Computational Creativity
City University, London

Formal frameworks for describing and
reasoning about creative behaviour
– Compare seach methods and outcomes
– Define value etc and reason about properties of
definitions

Pattern discovery and matching technogies
for multidimensional datasets
– Discover/locate geometrically identical structural
regions, possibly with gaps in multi-D data
– Example: 3D representations of atoms in space
for pharmacophore bonding models
University of A Corunha
Hybrid Society (HS)



Development framework
to validate and to allow 
the learning of various
computational models of
tasks which require

creativity and a social
behaviour
HS is based on machines
and humans living

together in a virtual and
“egalitarian” society
Solves the problem of
Value in a dynamic context.
Allows the comparison of
different computer
paradigms and systems.
Allow the collaboration
between humans and
computer systems
Allows the use of adaptive
techniques such us
Evolutionary Computation
and Artificial Neural
Networks
Creative Systems Group
University of Coimbra

Computational Models of Creativity
– Analogy
– Evolution
– Conceptual Blending
Models of Surprise
 Hybrid Societies for Creativity
Assessment

University of Edinburgh

Lakatos-style reasoning:
- Experts interact to build a common theory
- Counterexamples used to modify
conjectures; clarify concepts; improve
proofs
-
Ways of evaluating machine creativity
Universidad Complutense de Madrid

Ongoing research work:
– Knowledge intensive CBR
• CBRArm: framework for CBR + ontologies
– Generating narrative and metaphorical
texts, NLG architectures, CBR for text
generation
– CBR for Knowledge Management
• Java documentation, helpdesks
– Information Filtering + User Modeling
– Computer games
Creative Reasoning

Reasoning in non-standard ways to produce:
– “interesting”/valued/unexpected outputs
– emergent complex behaviour

Reconceptualise existing knowledge structures to get
new knowledge structures with added value
– using in a different way than they were intended
– lateral connections that weren’t there already

Heuristic reasoning
– Including sound and unsound methods

Post hoc verification
– value measurements for the domain are a pre-requisite
General Techniques
Conceptual blending
 Metaphorical/analogical reasoning
 Inductive inference
 Hypothesis repair
 Evolutionary methods

Conceptual Blending
Input: Atom (1)
E = electron
N = nucleus
r = rotates
around
E
Input: Solar System(2)
similar
o
r
N
P
S
similar
S much bigger
than P
N much bigger
than E
Electron = Atomic Planet
Nucleus = Atomic Sun
P = planet
S = sun
o = orbits around
E=P
BUT:
r=o
Electrons have a
statistical rather
than absolute position
in space
N=S
Gravity-like force
keeps the electrons
in orbit
about the nucleus Blend: Rutherford Atom
Metaphorical Reasoning
Entity
Abstraction
isa
isa
Abstraction
isa
Attribute
Object
isa
Relation
isa
isa
Substance
Quality
Communication
Poison (1)
Destructiveness
Disrespect
isa
Venom (1)
“A poison secreted by
certain animals”
isa
isa
Poison (2)
“Anything that harms or
destroys”
Insult (1)
“An artist who is master of
a particular style”
Inductive Inference

Predictive Induction
– Know the positives/negatives of a concept
– Search for a concept which fits categorisation
• Use examples as evidence for predictive accuracy
• Cross validate results

Descriptive Induction
– Search for rules which associate background
predicates, using data as empirical evidence
– (Sometimes) use deduction to prove rules found
Hypothesis Repair

Using a counterexample to repair a
faulty hypothesis by:
– Generalising from counterexample to a
property then stating the exception in the
hypothesis
– Generalising from the positives and then
limiting the hypothesis to these
Evolutionary Methods

Exploration of complex search spaces
– in non-uniform ways
– Based on biologically inspired evolutionary
notions such as gene recombination,
mutation, fitness functions
– Dynamically adaptive systems
Potential Applications

Levels of discovery
– You know what you are looking for,
• But you don’t know what it looks like
– You don’t know what you are looking for
• But you know you are looking for something
– You didn’t know you were even looking for
anything

Levels of search
– At the object level (millions/billions of data points)
– At the semantics level (tens of thousands of terms)
– At the meta-level (scores of techniques)
Possible (General) Application:
Ontology Maintenance

Ontologies standardise concepts
– And standardise relationships between them

Many areas see the need for ontologies
– Including scientific domains such as life sciences

Very important that the ontology represents
current scientific thinking
 Need to continually maintain ontology
– New nodes
– New links

Need to continually interpret ontology
– Large scale structures
Case Study – Gene Ontology

~14,000 terms from biology/genetics
– Process, function, structure
– Structured into hierarchies using isa/partof

Each term has genes associated
– ~ 1.3 million genes (from, e.g., GenBank)

Aims to unify biology
– Databases are in a bad state
• Different interpretations/notations/standards
Gene Ontology (Example)
Methods for Ontology
Maintenance

Mining rules between concepts using inductive
techniques (adds edges)
– Project to use HR for this in progress
– Project to use Progol to learn terminology

Conceptual blending
– Invent new concepts (nodes)

Metaphorical reasoning
– Look at structure to reorganise links

Hypothesis repair
– Explain genes which are seemingly misclassified
Proactive and Reactive
Applications

Proactive
– Attempt to make discoveries in GO
– Give value added when someone submits a
new term to the ontology

Reactive
– A new gene is added which (using sequence
alignments) is associated with “wrong” concept
– Creatively re-organise ontology to fix problem
The Bottom Line

We have solutions but not problems
– With respect to Life Sciences

Our application domains are disparate
– But our methods are general

We’re already thinking about certain
tasks/problems in life sciences
– Predictive toxicology
– Protein structure prediction

And we’re inventing our own problems

– Maintaining the Gene Ontology
But we really need to discuss what it is that standard
techniques do not yet give you
– And see what creative systems/techniques can do