Charles Taylor - Biology

1
Language and Biology Group
Charles Taylor - Biology Travis Collier Yoosook
Lee Yuan Yao Ed Stabler - Linguistics Greg
Kobele Jason Riggle
2
To/from humans
3
Explicit levelwell-formed, explicit models

• Are there or were there objects there?
• What kind were they?
• How many were there?
• What did they do?

4
Requirements at the Implicit Stage

• Robust
• changing environments/agents
• Wrong information
• noisy messages
• Adaptive
• unanticipated sources, events
• form new concepts
• different languages
• Self-configuring
• changing situations, goals

5
Outline

• Solution overview
• Partial solution - Evolving language
• Partial solution- Intrusion detection
• Formal Analysis
• Expressing knowledge with logic
• Creating and learning language syntax
• Semantics
• Grounding problem
• Passing D-structures

6
External World
Internal Representation Logical
Representation Decisions about what/whom to
communicate
Internal Representation Logical
Representation Decisions about what/whom to
communicate
English-like Language
Agent Language
English-like Language
Agent Language
Humans
Humans
7
Compression aids in generalization.
Compression distills experience into a schema or
model This compressed form can be succinct,
right, approximately correct or even wrong, but
it can be useful if it can be used to generalize
to situations different from previously
encountered. - Gell-Man
8
An example of compression
y mx b
9
Regular Language(Q, S , d , qo,F)

• Q set of all states (finite)
• S input alphabet (finite)
• qo initial state
• F set of final states
• d transition function (rewrite rules)
• (Q x d) ? Q

10
Example Rewrite Grammar
S
S ? NP VP NP ? D N NP ? D VP ? V VP ? V NP V ?
loves V ? eats D ? David D ? Mary N ? dog D ? the
NP VP
D V NP
Mary loves D
David
11
Minimum Description Length (MDL) Algorithm
- Rissanen Ristad
MDL-length grammar-encoding-length
data-encoding-length
Grammar-encoding-length (GEL) the cost of the
generalization Data-encoding-length (DEL) the
cost of the compression
12
Principle of Compression
Combine grammatical equivalents
S4
Jane
Amy
S5
S
S1
S2
Mary
likes
Caitlin
S6
Jane
Amy
S
S3
S1
S2
Caitlin
Mary
likes
13
3) Languages become more smooth?
14
Evolution of Language with semantics (Kirby)

• Loves (John, Mary) xxy
  zzy rrx
• Loves (Bill, Mary)
  xxy aab rrx
• Hits (Bill, John)
  mmn aab zzy
• aab Bill
• zzy John
• rrx Mary
• xxy Loves
• etc.

15
External World
Internal Representation Logical
Representation Decisions about what/whom to
communicate
Internal Representation Logical
Representation Decisions about what/whom to
communicate
English-like Language
Agent Language
English-like Language
Agent Language
Humans
Humans
16
Intrusion Detection Methods

• Specification-based methods
• ? ?(x)write(x, kernel)
• Pattern Matching
• signature of red code worm
• (could be specification-based - buffer overflow)
• Anomaly Detection
• Scan many ports in short time
• analogous to parts of our problem
• unanticipated changes in the system

17
Local Internal Events

• start (Subject Program EventNo Tstamp)
• chmod (Subject File Fpermissions EventNo Tstamp)
• open (Subject File Mode EventNo Tstamp)
• exec (Subject File Mode EventNo Tstamp)
• fork (Subject NewPID EventNo Tstamp)

18
(No Transcript)
19
(No Transcript)
20
Computer -linux
External World
1. Trace of activity
2. C objects - each file -each process
Internal Representation Logical
Representation Decisions about what/whom to
communicate
3. Prolog Environment - only interesting
parts, innate, human told, deduced
English-like Language
Humans