18
Associative Memory by Recurrent Neural Networks with Delay Elements Seiji MIYOSHI Hiro-Fumi YANAI Ma sato OKADA Kobe City College of Tech. Ibaraki Univ. RIK EN BSI , ERATO KDB JAPAN JAPAN JAPAN [email protected] www.kobe-kosen.ac.jp/ ~ miyoshi/

Associative Memory by Recurrent Neural Networks with Delay Elements Seiji MIYOSHI Hiro-Fumi YANAI Masato OKADA Kobe City College of Tech. Ibaraki Univ

Embed Size (px)

Citation preview

Page 1: Associative Memory by Recurrent Neural Networks with Delay Elements Seiji MIYOSHI Hiro-Fumi YANAI Masato OKADA Kobe City College of Tech. Ibaraki Univ

Associative Memory by Recurrent Neural Networks with Delay Elements

Seiji MIYOSHI Hiro-Fumi YANAI Masato OKADAKobe City College of Tech. Ibaraki Univ.   RIKEN BSI , ERATO KDB

JAPAN JAPAN JAPAN    

[email protected]

www.kobe-kosen.ac.jp/~miyoshi/

Page 2: Associative Memory by Recurrent Neural Networks with Delay Elements Seiji MIYOSHI Hiro-Fumi YANAI Masato OKADA Kobe City College of Tech. Ibaraki Univ

Background• Synapses of real neural systems seem to have delays.

• It is very important to analyze associative memory model with delayed synapses.

• Computer simulation is powerful method.

• Theoretical and analytical approach is indispensable to research on delayed networks.

There is a Limit on the number of neurons.However,

• Yanai-Kim theory by using Statistical Neurodynamics

Good Agreement with computer simulationComputational Complexity is O(L4t)

Simulating network with large delay steps is realistically impossible.

Page 3: Associative Memory by Recurrent Neural Networks with Delay Elements Seiji MIYOSHI Hiro-Fumi YANAI Masato OKADA Kobe City College of Tech. Ibaraki Univ

Objective

• To derive macroscopic steady state equations by using discrete Fourier transformation

• To discuss storage capacity quantitatively even for a large L limit (L: length of delay)

Page 4: Associative Memory by Recurrent Neural Networks with Delay Elements Seiji MIYOSHI Hiro-Fumi YANAI Masato OKADA Kobe City College of Tech. Ibaraki Univ

Recurrent Neural Network with Delay Elements

Delay E lement1 l L-1

J

J

ij ij ij

ji ji ji ji

ijL-1

L-1

l

l

1

1

0

0

J

J

J

J

J

J

Neuron

i

N

1

j

Model

Page 5: Associative Memory by Recurrent Neural Networks with Delay Elements Seiji MIYOSHI Hiro-Fumi YANAI Masato OKADA Kobe City College of Tech. Ibaraki Univ

• Overlap

Model

• Discrete Synchronous Updating Rule

• Correlation Learning for Sequence Processing

Page 6: Associative Memory by Recurrent Neural Networks with Delay Elements Seiji MIYOSHI Hiro-Fumi YANAI Masato OKADA Kobe City College of Tech. Ibaraki Univ

Macrodynamical Equations by Statistical NeurodynamicsYanai & Kim(1995)   Miyoshi, Yanai & Okada(2002)

Page 7: Associative Memory by Recurrent Neural Networks with Delay Elements Seiji MIYOSHI Hiro-Fumi YANAI Masato OKADA Kobe City College of Tech. Ibaraki Univ

Initial Condition of the NetworkOne Step Set Initial Condition• Only the states of neurons are

set explicitly.

• The states of delay elements are set to be zero.

All Steps Set Initial Condition• The states of all neurons and all

delay elements are set to be close to the stored pattern sequences.

• If they are set to be the stored pattern sequences themselves

≡ Optimum Initial Condition

NeuronDelay E lement

i

N

1

1 l L-1

j

J

J

ij ij ij

ji ji ji ji

ijL-1

L-1

l

l

1

1

0

0

J

J

J

J

J

J

set zero

NeuronDelay E lement

i

N

1

1 l L-1

j

J

J

ij ij ij

ji ji ji ji

ijL-1

L-1

l

l

1

1

0

0

J

J

J

J

J

J

set

Page 8: Associative Memory by Recurrent Neural Networks with Delay Elements Seiji MIYOSHI Hiro-Fumi YANAI Masato OKADA Kobe City College of Tech. Ibaraki Univ

Dynamical Behaviors of Recall Processo

ver

lap

m

1

0.8

0.6

0.4

0.2

0

time step t

0 5 10 15 20 25 30

time step t

0 5 10 15 20 25 30

ov

erla

p

m

1

0.8

0.6

0.4

0.2

0

All Steps Set Intial Condition Loading rateα=0.5

Length of delay L=3

Simulation( N=2000)

Theory

Page 9: Associative Memory by Recurrent Neural Networks with Delay Elements Seiji MIYOSHI Hiro-Fumi YANAI Masato OKADA Kobe City College of Tech. Ibaraki Univ

time step t

ov

erla

p

m

1

0.8

0.6

0.4

0.2

0

0 5 10 15 20 25 30

time step t

ov

erla

p

m

1

0.8

0.6

0.4

0.2

0

0 5 10 15 20 25 30

Dynamical Behaviors of Recall Process

  All Steps Set Intial Condition Loading rateα=0.5

Length of delay L=2

Simulation( N=2000)

Theory

Page 10: Associative Memory by Recurrent Neural Networks with Delay Elements Seiji MIYOSHI Hiro-Fumi YANAI Masato OKADA Kobe City College of Tech. Ibaraki Univ

Loading rates α - Steady State Overlaps m

Simulation( N=500)

Ov

erla

p

m

1

0

0.2

0

L=1 L=3

One Step Set Optimum

L=10

L=10

0.5 1 1.5 2

0.4

0.6

0.8

αLoading RateO

ver

lap

m

1

0

0.2

0.4

0.6

0.8

0 0.5 1 1.5 2

αLoading Rate

L=3

One Step Set Optimum

L=10L=10L=1

Theory

Page 11: Associative Memory by Recurrent Neural Networks with Delay Elements Seiji MIYOSHI Hiro-Fumi YANAI Masato OKADA Kobe City College of Tech. Ibaraki Univ

Length of delay L - Critical Loading Rate αC

Cri

tica

l Lo

ad

ing

Ra

te

1

2

0.21 2 3 4 5 6 7 8 9 10

2.2

1.2

0.4

1.4

0.6

1.6

0.8

1.8

Length of Delay L

Optimum

One Step Set

Page 12: Associative Memory by Recurrent Neural Networks with Delay Elements Seiji MIYOSHI Hiro-Fumi YANAI Masato OKADA Kobe City College of Tech. Ibaraki Univ

Macrodynamical Equations by Statistical NeurodynamicsYanai & Kim(1995)   Miyoshi, Yanai & Okada(2002)

• Good Agreement with Computer Simulation

• Computational Complexity is O(L4t)

Page 13: Associative Memory by Recurrent Neural Networks with Delay Elements Seiji MIYOSHI Hiro-Fumi YANAI Masato OKADA Kobe City College of Tech. Ibaraki Univ

Macroscopic Steady State Equations• Accounting for Steady State• Parallel Symmetry in terms of Time Steps• Discrete Fourier Transformation

Page 14: Associative Memory by Recurrent Neural Networks with Delay Elements Seiji MIYOSHI Hiro-Fumi YANAI Masato OKADA Kobe City College of Tech. Ibaraki Univ

Loading rates α - Steady State Overlaps m

ov

erla

p

m

1

0.8

0.6

0.4

0.2

0

0.1 1 10 100

L=1

1000 10000

αLoading Rate

ov

erla

p

m

1

0.8

0.6

0.4

0.2

0

0.1 1 10 100

L= 1L= 2

1000 10000

αLoading Rate

ov

erla

p

m

1

0.8

0.6

0.4

0.2

0

0.1 1 10 100

L= 1L= 2L= 3

1000 10000

αLoading Rate

ov

erla

p

m

1

0.8

0.6

0.4

0.2

0

0.1 1 10 100

L= 1L= 2L= 3

L= 10

1000 10000

αLoading Rate

ov

erla

p

m

1

0.8

0.6

0.4

0.2

0

0.1 1 10 100

L= 1L= 2L= 3

L= 10L=100

1000 10000

αLoading Rate

ov

erla

p

m

1

0.8

0.6

0.4

0.2

0

0.1 1 10 100

L= 1L= 2L= 3

L= 10L=100

L= 1000

1000 10000

αLoading Rate

ov

erla

p

m

1

0.8

0.6

0.4

0.2

0

0.1 1 10 100

L= 1L= 2L= 3

L= 10L=100

L= 1000L= 10000

1000 10000

αLoading Rate

ov

erla

p

m

1

0.8

0.6

0.4

0.2

0

0.1 1 10 100

L= 1L= 2L= 3

L= 10L=100

L= 1000L= 10000

L=100000

1000 10000

αLoading Rate

Page 15: Associative Memory by Recurrent Neural Networks with Delay Elements Seiji MIYOSHI Hiro-Fumi YANAI Masato OKADA Kobe City College of Tech. Ibaraki Univ

Loading rates α - Steady State Overlaps m

Simulation( N=500)

Ov

erla

p

m

1

0

0.2

0

L=1 L=3

One Step Set Optimum

L=10

L=10

0.5 1 1.5 2

0.4

0.6

0.8

αLoading RateO

ver

lap

m

1

0

0.2

0.4

0.6

0.8

0 0.5 1 1.5 2

αLoading Rate

L=3

One Step Set Optimum

L=10L=10L=1

Theory

Page 16: Associative Memory by Recurrent Neural Networks with Delay Elements Seiji MIYOSHI Hiro-Fumi YANAI Masato OKADA Kobe City College of Tech. Ibaraki Univ

Loading rate α - Steady State Overlap

ov

erla

p

m

1

0.8

0.6

0.4

0.2

0

0.1 1 10 100

L= 1L= 2L= 3

L= 10L=100

L= 1000L= 10000

L=100000

1000 10000

αLoading Rate

Page 17: Associative Memory by Recurrent Neural Networks with Delay Elements Seiji MIYOSHI Hiro-Fumi YANAI Masato OKADA Kobe City College of Tech. Ibaraki Univ

Storage Capacity of Delayed Network

Sto

rag

e C

ap

ac

ity

αC

Number of Delays L

1

10.1

10

10

100

100

1000

1000

10000

10000 100000

100000

Storage Capacity = 0.195 L

Page 18: Associative Memory by Recurrent Neural Networks with Delay Elements Seiji MIYOSHI Hiro-Fumi YANAI Masato OKADA Kobe City College of Tech. Ibaraki Univ

Conclusions• Yanai-Kim theory (macrodynamical equations for del

ayed network) is re-derived.

• Steady state equations are derived by using discrete Fourier transformation.

• Storage capacity is 0.195 L in a large L limit.

→ Computational Complexity is O(L4t)

→ Intractable to discuss macroscopic properties in a large L limit

→ Computational complexity does not formally depend on L→ Phase transition points agree with those under the optimum initial conditions, that is, the Storage Capacities !