• Keine Ergebnisse gefunden

„Algebraic Mind“, Marcus

N/A
N/A
Protected

Academic year: 2021

Aktie "„Algebraic Mind“, Marcus"

Copied!
24
0
0

Wird geladen.... (Jetzt Volltext ansehen)

Volltext

(1)

Where does the Machinery of Symbol Manipulation come from

[Seminar: Cognitive Architectures]

Chapter 6

„Algebraic Mind“, Marcus

Anne Lämmel, Jiayong Liu, Michael Elbers

14.05.2003

(2)

Overview

Is the Machinery of Symbol Manipulation (MSM) innate ?

Is Symbol Manipulation (SM) adaptive ?

(evolutionary aspects)

How can SM grow ?

(developmental aspects)

(3)

Innateness Hypothesis

• MSM belongs to the set of things available to children prior to experience

• MSM consists of innately given representational formats for:

– allowing OPOVARs

• set of operations computable over variables

• apparatus for combining operations – structured combinations

– the distinction of individuals from kinds

(4)

The Learnability Argument

• Mostly in context of language acquisition

Plural Compounds (Gordon, 1985)

Infant Concept of Object (Spelke, 1994)

• Something has to be innate

• No absolute proof for innateness yet,

but a challenge to theories of learning

that depend on much experience

(5)

Overview

Is the Machinery of Symbol Manipulation (MSM) innate ?

Is Symbol Manipulation (SM) adaptive ?

(evolutionary aspects)

How can SM grow ?

(developmental aspects)

(6)

Evolutionary Aspects

• Has Natural Selection shaped the mind and thus the MSM ?

• Many conflicting views on Natural Selection

• Marcus expects evidence from genetics (Phylogenetic History)

• Marcus discusses evidence for SM in animals

– How pervasive is the MSM in the animal kingdom ?

– How may SM have been advantageous in evolution ?

(7)

Symbols

• No consensus on a definition for „symbol“

• Marcus focusses on Equivalence Classes

Frog Example (Lettvin et al., 1959) – Three kinds of Equivalence Classes :

• perceptual features

• non-perceptual features

(e.g. convention, organism‘s goals, knowledge of physical world) – Sea Lion (Schusterman & Kastak, 1998)

(8)

Rules

“Presumably, the ability to learn new rules depends on prior

selection for mechanisms for representing and generalizing rules.“

• Two aspects to rules:

– representation & generalization (pervasive) – learning new rules (not so pervasive)

• Supporting experiments:

Match-to-Sample Tasks (primates, parrots, sea lions)

(Solar) Azimuth Function (desert ants & honey bees)

(9)

Rules: Azimuth Experiments

• Azimuth Function (for Navigation) :

calculating the sun‘s position depending on latitude and season – implausibility of look-up table of values

Limited Exposure (Lindauer, 1959)

– implausibility of extensive sampling/experience

Learning Azimuth with MLP (Dickinson & Dyer, 1996)

– MLP only operates/generalizes correctly on trained data – Assumption: animals estimate the azimuth by using

recent position data and the time elapsed

=> rule operating over numeric variables

(10)

Rules (concluded)

• Is the ability to manipulate numeric variables

more pervasive than to manipulate non-numeric variables ?

– maybe the non-numerics only evolved with social exchange (exception: mimicking songbirds ?)

• Maybe only humans can adapt domain-specific machinery

for many more general purposes ?

(11)

Structured Representations

• Probably unique to humans to have structured (hierarchical) representations (SR)

– some animals may learn to some extent, but no

evidence for representation of hierarchical structure (basic word-order, temporal orderings)

• Two theories of the origin of SR :

– Communication / Language

– other aspects like Motor control / Planning

But: we need to better understand the neural substrate

and the role of genes first

(12)

Structured Representations

• Marcus‘ speculative account of evolutionary advantage of using SR in Language

before language: unique internal identifiers

then: combining sound elements (words) unstructuredly

(e.g. if the thing referred to is not in plain sight)

then: simple structured combinations

(e.g. „Johnny give“ vs. „give Johnny“)

finally: application to nonlinguistic cognition

(13)

Individuals

• Adaptive Advantages of Tracking Individuals :

– hunting

– food sources (object permanence) – family members

– social exchange

Chicks (Regolin et al., 1995)

– innate object permanence ?

(14)

Overview

Is the Machinery of Symbol Manipulation (MSM) innate ?

Is Symbol Manipulation (SM) adaptive ?

(evolutionary aspects)

How can SM grow ?

(developmental aspects)

(15)

Developmental Aspects

• DNA as Blueprint

– probably works for simple brains, but not for the

complex structure of human brain

(16)

• the Brain is more complex than the Genome

• Variability of the Brain

• Flexibility of the Brain (Elman, 1996):

– size dependence – rewiring

– transplants

– recovery from damage

(17)

Experiments by Elman et al.

• Argument against Nativism (Brain Flexibility)

• Learning is the major Factor

– „attention grabbers“ for fast learning

• architectual and timing differences play a minor

role

(18)

MLP-Models by Elman et al.

• Counterarguments by Marcus:

– input nodes serve as innate representations – can not escape the DNA as Blueprint View

model relies on a lot of prespecification

– can not capture all aspects of Human Cognition

! How we generalize new items

! How we represent structured knowledge

! How we represent and track individual over time

(19)

Strength of Evidence for Flexibility

• Limits on Size Dependence

• Limits on Rewiring

• Limits on Transplants

• Limits on Recovery from Brain Damage

• Constrained Flexibility rather than cortical

Equipotentiality

(20)

Is Learning essential for Brain Development?

• important Aspects of Brain Structure are organized prior to experience

Ocular dominance columns (Katz & Shatz, 1996)

(21)

Resolving an Apparent Paradox

• Learning vs Blueprint

– (Elman et al. vs. Katz & Shatz)

• DNA as recipe: Structure=Cascade+Signaling

(22)

Proposal:

The conjunction of cell-to-cell signaling and

cascading genes provides a good account of how

nonmental aspects of embryos develop, and Marcus argues that comparable mechanisms plan an

important role in the development of the brainMSM could be constructed by genetically driven

Mechanisms and in tandem with activity

dependence.

(23)

Summary

Is the Machinery of Symbol Manipulation (MSM) innate ?

Is Symbol Manipulation (SM) adaptive ?

(evolutionary aspects)

How can SM grow ?

(developmental aspects)

(24)

Summary

• MSM is innate, but no final proof yet

• Pervasiveness of MSM (evolutionary aspect)

• DNA as recipe (developmental aspect)

Referenzen

ÄHNLICHE DOKUMENTE

A good reference to illustrate the point is eAQUA, an interdisciplinary project set up between the departments for classical studies at the University of Leipzig, Heidelberg,

Complete these sentences with a past suggestion using the verb in brackets.. Nick was

Complete these sentences with a past suggestion using the verb in brackets. Nick was

My English cousin __doesn’t have to____ do military service – it isn't compulsory there.. You __mustn’t___ use a dictionary during the exam – it's against

Secondly suppose that a group of workers (say group i) is more productive than another (say group -i); for instance, one may imagine that group i is represented by high-skilled

Then in Chapter 3 we deal with the second approach applied to stochastic optimal control problems, showing how the solution to the Backward Stochastic Differential Equation

Note that under the decision to supply for the system Si the proper W R Y we can divide this total inflow xi in the corresponding components xik independently of other systems S

Third, financial development has a large causal effect in the reduction of macroeconomic volatility as a result of liquidity provision by the financial sector (Raddatz,