96
The Neuronal Replicator Hypothesis Chrisantha Fernando & Eors Szathmary CUNY, December 2009 1 Collegium Budapest (Institute for Advanced Study), Budapest, Hungary 2 Centre for Computational Neuroscience and Robotics, Sussex University, UK 3 MRC National Institute for Medical Research, Mill Hill, London, UK 4 Parmenides Foundation, Kardinal-Faulhaber-Strase 14a, D-80333 Munich, Germany 5 Institute of Biology, Eötvös University, Pázmány Péter sétány 1/c, H-1117 Budapest, Hungary

The Neuronal Replicator Hypothesis

  • Upload
    carl

  • View
    61

  • Download
    4

Embed Size (px)

DESCRIPTION

The Neuronal Replicator Hypothesis. Chrisantha Fernando & Eors Szathmary CUNY, December 2009 1 Collegium Budapest (Institute for Advanced Study), Budapest, Hungary 2 Centre for Computational Neuroscience and Robotics, Sussex University, UK - PowerPoint PPT Presentation

Citation preview

Page 1: The Neuronal Replicator Hypothesis

The Neuronal Replicator Hypothesis

Chrisantha Fernando & Eors Szathmary

CUNY, December 2009

1Collegium Budapest (Institute for Advanced Study), Budapest, Hungary2Centre for Computational Neuroscience and Robotics, Sussex University, UK

3MRC National Institute for Medical Research, Mill Hill, London, UK4Parmenides Foundation, Kardinal-Faulhaber-Strase 14a, D-80333 Munich, Germany

5Institute of Biology, Eötvös University, Pázmány Péter sétány 1/c, H-1117 Budapest, Hungary

Page 2: The Neuronal Replicator Hypothesis

Visiting Fellow MRC National Institute for

Medical ResearchLondon

Post-DocCenter for Computational

Neuroscience and RoboticsSussex University

Marie Curie FellowCollegium Budapest

(Institute for Advanced Study)Hungary

Page 3: The Neuronal Replicator Hypothesis

The Hypothesis•Evolution by natural selection takes

place in the brain at rapid timescales and contributes to solving cognitive/behavioural search problems.

•Our background is in evolutionary biology/the origin of non-enzymatic template replication/evolutionary robotics/computational neuroscience.

Page 4: The Neuronal Replicator Hypothesis

Outline

•Limitations of some proposed search algorithms, e.g.

•Reward biased stochastic search

•Reinforcement Learning

•How copying/replication of neuronal data structures can alleviate these limitations.

Page 5: The Neuronal Replicator Hypothesis

•Mechanisms of neuronal replication

•Applications and future work

Page 6: The Neuronal Replicator Hypothesis

Simple Search Tasks

•Behavioural and neuropsychological learning tasks can be solved by stochastic-hill climbing

•Stroop Task

•Wisconsin Card Sorting Task (WCST)

•Instrumental Conditioning in Spiking Neural Networks

•Simple inverse kinematics problem

Page 7: The Neuronal Replicator Hypothesis

Stochastic Hill-Climbing

• Initially P(xi = 1) = 0.5, Initial reward = 0

• Make random change to P

• Generate M examples of binary strings

• Calculate reward

• If r(t) > r(t-1), keep changes of P, else revert to previous P values.

• One solution, change solution, keep good changes, loose bad changes.

0.5 0.5 0.5 0.5 0.5

0.5 0.5 0.5 0.5 0.5

0.8 0.5 0.5 0.4 0.5

Page 8: The Neuronal Replicator Hypothesis

Can get stuck on local optima

Page 9: The Neuronal Replicator Hypothesis

Stroop TaskGreen Red Blue Purple Blue Purple

Blue Purple Red Green Purple Green

Name the colour of the words.

Page 10: The Neuronal Replicator Hypothesis

Dehaene et al, 1998

dW = Reward x pre x postDecreased reward -> Instability in workspace

Page 11: The Neuronal Replicator Hypothesis

WCST•Each card has several “features”.

Subjects must sort cards according to a feature (color, number, shape, size).

Page 12: The Neuronal Replicator Hypothesis

•Rougier et al 2005. PFC weights stabilised if expected reward obtained, destabilised if expected reward not obtained, i.e. TD learning

Page 13: The Neuronal Replicator Hypothesis

Instrumental Conditioning

Page 14: The Neuronal Replicator Hypothesis

In a spiking neural net

Izhikevich 2007

• Simple spiking model • Random connections • STDP • Delayed reward • Eligibility traces• Synapse selected

Page 15: The Neuronal Replicator Hypothesis

• Simple spiking model

Page 16: The Neuronal Replicator Hypothesis

STDP

Page 17: The Neuronal Replicator Hypothesis
Page 18: The Neuronal Replicator Hypothesis
Page 19: The Neuronal Replicator Hypothesis

Time tpre

Page 20: The Neuronal Replicator Hypothesis

Time tpost

Page 21: The Neuronal Replicator Hypothesis

Interval = Tpost - Tpre

Page 22: The Neuronal Replicator Hypothesis

Time tpost

Page 23: The Neuronal Replicator Hypothesis
Page 24: The Neuronal Replicator Hypothesis
Page 25: The Neuronal Replicator Hypothesis
Page 26: The Neuronal Replicator Hypothesis

Time tpre

Page 27: The Neuronal Replicator Hypothesis

Interval = Tpost - Tpre

Page 28: The Neuronal Replicator Hypothesis
Page 29: The Neuronal Replicator Hypothesis

A simple 2D inverse kinematics

problem

Page 30: The Neuronal Replicator Hypothesis

Reinforcement Learning

• For large problems a tabular representation of state-action pairs is not possible.

• How does compression of state representation occur? Function approximation

• Domain-specific knowledge provided by the designer, e.g. TD-Gammon was dependent on Tesauro’s skillful design of a non-linear multilayered neural network, used for value function approximation in the Backgammon domain consisting of approximately 1020 states” p20 [51].

Page 31: The Neuronal Replicator Hypothesis

So far…•SHC works on simple problems

•RL is a sophisticated kind of SHC

•In order for RL/SHC to work, action/value representations must fit the problem domain.

•RL doesn’t explain how appropriate data-structures/representations arise.

Page 32: The Neuronal Replicator Hypothesis

Large search space sorandom search or

exhaustive search not possible.

Representation criticallocal optima.

Requires internal sub-goals, no explicit

reward.

What neural mechanisms underlie complex search?

Page 33: The Neuronal Replicator Hypothesis

What is natural selection?

Some hereditary traits affect survival and/or fertility

1. multiplication

2. heredity

3. variability

Page 34: The Neuronal Replicator Hypothesis

Natural selection reinvented itself

Page 35: The Neuronal Replicator Hypothesis

Evolutionary Computation•Solving problems by EC also

requires decisions about genetic representations

•And about fitness functions

•For example, we use EC to solve the 10 coins problem

Page 36: The Neuronal Replicator Hypothesis
Page 37: The Neuronal Replicator Hypothesis

Fitness function•Convolution of desired inverted

triangle over grid

•Instant fitness = number of coins occupying he inverted triangle template

•An important question is how such fitness functions (subgoals/goals) could themselves be bootstrapped in cognition.

Page 38: The Neuronal Replicator Hypothesis
Page 39: The Neuronal Replicator Hypothesis

Michael Ollinger, Parmenides Foundation, Munich

Page 40: The Neuronal Replicator Hypothesis

Structuring Phenotypic Variation

•Natural Selection can act on

•genetic representations

•variability properties (genetic operators, e.g mutation rates)

Page 41: The Neuronal Replicator Hypothesis

A

Variation in Variability

Improvement of representations for free…

Page 42: The Neuronal Replicator Hypothesis

B

Page 43: The Neuronal Replicator Hypothesis

Non-trivial Neutrality

g1

g2p

ed 1

ed 2

Adapted from Toussaint 2003

Page 44: The Neuronal Replicator Hypothesis

Population Search•Natural selection allows

redistribution of search resources between multiple solutions.

•We propose that multiple (possibly interacting) solutions to a search problem exist at the same time in the neuronal substrate.

Page 45: The Neuronal Replicator Hypothesis

AAB

C

D

AAB

C

D

A B C D

A B C D

Page 46: The Neuronal Replicator Hypothesis

A B C D

AD’D’’

D’’’D

C

DA

B

A B C D

AAB

C

D

A B C D

D’ D’’ D’’’ D

Waste

Page 47: The Neuronal Replicator Hypothesis

Can units of selection exist in the

brain?•We propose 3 possible mechanisms

•Copying of connectivity patterns

•Copying of bistable activity patterns

•Copying of spatio-temporal spike patterns & explicit rules

Page 48: The Neuronal Replicator Hypothesis

Copying of connectivity

patterns

Page 49: The Neuronal Replicator Hypothesis

How to copy small neuronal circuits

DNA neuronal network

Page 50: The Neuronal Replicator Hypothesis

STDP and causal inference

Page 51: The Neuronal Replicator Hypothesis
Page 52: The Neuronal Replicator Hypothesis

With error correction and sparse activation

Page 53: The Neuronal Replicator Hypothesis

1 + 1 Evolution Stratergy

Page 54: The Neuronal Replicator Hypothesis

Copying of bistable activity patterns

Page 55: The Neuronal Replicator Hypothesis
Page 56: The Neuronal Replicator Hypothesis

1 bit copy

Page 57: The Neuronal Replicator Hypothesis

Hebbian Learning can Structure Exploration Distributions

Page 58: The Neuronal Replicator Hypothesis
Page 59: The Neuronal Replicator Hypothesis
Page 60: The Neuronal Replicator Hypothesis
Page 61: The Neuronal Replicator Hypothesis
Page 62: The Neuronal Replicator Hypothesis
Page 63: The Neuronal Replicator Hypothesis
Page 64: The Neuronal Replicator Hypothesis
Page 65: The Neuronal Replicator Hypothesis
Page 66: The Neuronal Replicator Hypothesis

- Search in biased towards previous local optima

Page 67: The Neuronal Replicator Hypothesis
Page 68: The Neuronal Replicator Hypothesis
Page 69: The Neuronal Replicator Hypothesis
Page 70: The Neuronal Replicator Hypothesis

The Origin of Heredity in Neuronal Networks.

Phenotype 2

Phenotype 1

M2

M1

C

Genotype 1

Genotype 2

CM2= M1

C = M2-1M1

Page 71: The Neuronal Replicator Hypothesis

Non-local, e.g. requires ATA

Page 72: The Neuronal Replicator Hypothesis

Stochastic hill climbing can select for neuronal template replication

M2

M1

C

Genotype 1

Genotype 2

EEErrorError

Page 73: The Neuronal Replicator Hypothesis
Page 74: The Neuronal Replicator Hypothesis

Copying of Spatiotemporal

Spike Patterns & Explicit Rules

Page 75: The Neuronal Replicator Hypothesis

Spatiotemporal spike patterns

Page 76: The Neuronal Replicator Hypothesis
Page 77: The Neuronal Replicator Hypothesis
Page 78: The Neuronal Replicator Hypothesis
Page 79: The Neuronal Replicator Hypothesis
Page 80: The Neuronal Replicator Hypothesis

ABA vs ABB

DD vs DS

Visual shift-invariancemechanisms applied

to linguistics.

Page 81: The Neuronal Replicator Hypothesis

APPLICATIONS

Page 82: The Neuronal Replicator Hypothesis

•Evolution of Predictors (Feed-forward Models/Emulators/Bayesian Causal Networks).

•First derivative of predictability

•Evolution of Linguistic Construction

•Evolution of controllers for robot hand-manipulation

•Evolution of Productions in ACT-R/Copycat

•Evolution of representations and search for insight problem solving.

Page 83: The Neuronal Replicator Hypothesis
Page 84: The Neuronal Replicator Hypothesis

Operations to construct a BN

Larranaga et al, 1996. Structure Learning of Bayesian Networks by Genetic Algorithms.Kemp & Tenenbaum, 2008. The discovery of structural form.

Page 85: The Neuronal Replicator Hypothesis

Luc Steels et al, Sony Labs

Page 86: The Neuronal Replicator Hypothesis

Istvan Zacher Collegium Budapest (Institute for Advanced Study)

Page 87: The Neuronal Replicator Hypothesis

K(v)

S(p) C(p)0 1

Page 88: The Neuronal Replicator Hypothesis

K(v)

S(p) C(p)0 1

Rules

Page 89: The Neuronal Replicator Hypothesis

K(v)

S(p) C(p)0 1

Rules

Page 90: The Neuronal Replicator Hypothesis

K(v)

S(p) C(p)0 1

Rules

Page 91: The Neuronal Replicator Hypothesis

K(v)

S(p) C(p)0 1

Rules

KC

Page 92: The Neuronal Replicator Hypothesis

K(v)

S(p) C(p)0 1

Rules

KC S

Page 93: The Neuronal Replicator Hypothesis

Rules

KC S

Page 94: The Neuronal Replicator Hypothesis

Rules

KC S

K(v)

S(p) C(p)0 1

Page 95: The Neuronal Replicator Hypothesis

Helge Ritter, Bielefeld, Germany

Page 96: The Neuronal Replicator Hypothesis

Thanks toRichard GoldsteinRichard Watson

Dan BushEugine Izhikevich

Phil HusbandsLuc Steels

K.K. KarishmaAnna Fedor, Zoltan Szatmary, Szabolcs Szamado, Istvan Zachar

Anil Seth