neural network with memory and cognitive functions janusz a. starzyk, and yue li school of...

28
Neural Network with Memory and Cognitive Functions Janusz A. Starzyk, and Yue Li School of Electrical Engineering and Computer Science Ohio University, Athens, OH 45701, U.S.A. David D. Vogel Ross University School of Medicine, Commonwealth of Dom inica

Upload: thomas-turner

Post on 12-Jan-2016

217 views

Category:

Documents


1 download

TRANSCRIPT

Page 1: Neural Network with Memory and Cognitive Functions Janusz A. Starzyk, and Yue Li School of Electrical Engineering and Computer Science Ohio University,

Neural Network with Memory and Cognitive Functions

Janusz A. Starzyk, and Yue LiSchool of Electrical Engineering and Computer Science

Ohio University, Athens, OH 45701, U.S.A.

David D. VogelRoss University School of Medicine, Commonwealth of Dominica

Page 2: Neural Network with Memory and Cognitive Functions Janusz A. Starzyk, and Yue Li School of Electrical Engineering and Computer Science Ohio University,

2

OrganizationOrganization

• Motivation

• R-nets

• Statistical model of R-nets

• Simulation results of statistical model

• Simulation results of R-nets

• Conclusions

Page 3: Neural Network with Memory and Cognitive Functions Janusz A. Starzyk, and Yue Li School of Electrical Engineering and Computer Science Ohio University,

3

• Perception

• Capacity to learn and memorize useful things

• Spatio-temporal memories

• Motor skills in relation to sensing and anticipation

• Surviving in complex environment and adaptation

• Value driven behavior (formulation of goals and their implementation)

• Abstract thinking and action planning

• Ability to communicate

• Intuition and creativity

• Consciousness

MotivationMotivation

To design an intelligent machine capable of

Page 4: Neural Network with Memory and Cognitive Functions Janusz A. Starzyk, and Yue Li School of Electrical Engineering and Computer Science Ohio University,

4

Problems of Classical AIProblems of Classical AI

• Lack of robustness and generalization

• No real-time processing

• Central processing of information by a single processor

• No natural interface to environment

• No self-organization

• Need to write software

Page 5: Neural Network with Memory and Cognitive Functions Janusz A. Starzyk, and Yue Li School of Electrical Engineering and Computer Science Ohio University,

5

Intelligent BehaviorIntelligent Behavior

• Emergent from interaction with environment

• Based on large number of sparsely connected neurons

• Asynchronous

• Interact with environment through sensory-motor system

• Value driven

• Adaptive

• Coordinated through agent’s

– multiple sensory-motor modalities

– constraints from morphology and materials

– generation of correlations through physical process

– basis for cross-modal associationsFrom http://tokyolectures.org/

Page 6: Neural Network with Memory and Cognitive Functions Janusz A. Starzyk, and Yue Li School of Electrical Engineering and Computer Science Ohio University,

6

Sparse Connectivity

The brain is sparsely connected. (Unlike most neural nets.)

A neuron in cortex may have on the order of 100,000 synapses.

There are more than 1010 neurons in the brain. Fractional connectivity is very low: 0.001%.

Implications: 

• Connections are expensive biologically since they take up space, use energy, and are hard to wire up correctly.

• Therefore, connections are valuable.

• The pattern of connection is under tight control.

• Short local connections are cheaper than long ones.

Our model makes extensive use of local connections.

Page 7: Neural Network with Memory and Cognitive Functions Janusz A. Starzyk, and Yue Li School of Electrical Engineering and Computer Science Ohio University,

7Human Brain

at Birth 6 Years Old 14 Years Old

Neuron Structure and Self-Organizing Principles

Page 8: Neural Network with Memory and Cognitive Functions Janusz A. Starzyk, and Yue Li School of Electrical Engineering and Computer Science Ohio University,

8

“The basic unit of cortical operation is the minicolumn … It contains of the order of 80-100 neurons except in the primate striate cortex, where the number is more than doubled.

The minicolumn measures of the order of 40-50 m in transverse diameter, separated from adjacent minicolumns by vertical, cell-sparse zones … “

(VB Mountcastle (2003). Introduction [to a special issue of Cerebral Cortex on columns]. Cerebral Cortex, 13, p. 2)

Stain of cortex in planum temporale.

Cortical Minicolumns

Page 9: Neural Network with Memory and Cognitive Functions Janusz A. Starzyk, and Yue Li School of Electrical Engineering and Computer Science Ohio University,

9

R-netsR-nets

• R-nets are defined as randomly connected artificial neural networks with primary and secondary neurons

• They implement distributed memories able to recall input patterns

• R-net with 106 excitatory neurons and brain-like connectivity will store at least 2x108 bits of information

• R-nets have been used to construct networks with higher cognitive functions.

• A statistical model of R-nets is presented and results are compared with simulated R-nets.

Page 10: Neural Network with Memory and Cognitive Functions Janusz A. Starzyk, and Yue Li School of Electrical Engineering and Computer Science Ohio University,

10

Properties of R-netsProperties of R-nets

• Biological plausibility• Similar to Hebbian networ

ks• Sparsly connected • Large storage capacities • Use inhibition to prevent n

eurons not associated with a recalled pattern from firing

• Store sequential patterns and attributes

From Marr, D. (1971). Simple memory: a theory for archicortex. Philosophical Transaction of the Royal Society of London, B. 262, 23‑81.

Page 11: Neural Network with Memory and Cognitive Functions Janusz A. Starzyk, and Yue Li School of Electrical Engineering and Computer Science Ohio University,

11

R-netsR-nets

• During training, an input pattern is presented to the R-net by activating a selected cluster C of primary neurons.

• All links between active primary neurons are trained.

• During recall a subset of one of the stored patterns is presented to the input, activating corresponding primary neurons (initial recall set).

• links from initial primary neuron 1 to destination primary neurons 1 and 2 are untrained. So is the link from initial primary neuron 2 to destination primary neuron 1.

• The only trained link is from initial primary neuron 2 to destination primary neuron 2.

• Both secondary neurons are strongly activated.

Page 12: Neural Network with Memory and Cognitive Functions Janusz A. Starzyk, and Yue Li School of Electrical Engineering and Computer Science Ohio University,

12

R-netsR-nets

• During recall, secondary neurons sum the weighted projections of active primary neurons

, , ,i x i e e xa W a

ai,x represents the activity if the ith secondary neuron in the xth cycle;ae,x is the current activity of the eth primary neuron with possible values 0 or 1; Wi,e is the weight of the projection of the eth primary neuron onto the ith secondary neuron with possible values of 1 (untrained) or 10 (trained).

Page 13: Neural Network with Memory and Cognitive Functions Janusz A. Starzyk, and Yue Li School of Electrical Engineering and Computer Science Ohio University,

13

R-netsR-nets

• The primary neurons are synchronously updated.

, , 1 , 1 1e x e x i xI I I

Ie,x is the inhibition of the eth primary neuron on the xth cycle,

Ii,x-1 is a function of the ith secondary neuron, which is given by:

, 1 , 1i x i xI a

, 1 1i xI

, 1 0i xI

, 1 10i xa

, 1 10i xa

, 1 10i xa

and the projection of the ith secondary neuron is untrained

and the projection of the ith secondary neuron is trained

if

if

if

Page 14: Neural Network with Memory and Cognitive Functions Janusz A. Starzyk, and Yue Li School of Electrical Engineering and Computer Science Ohio University,

14

Statistical Model of R-nets Statistical Model of R-nets

P-primary neurons S-secondary neuronsKp-primary neurons’ outgoing sets Ks-secondary neurons’ outgoing sets

sp kSkP **

P

PciSet of primary neurons linked to ci ciP

Page 15: Neural Network with Memory and Cognitive Functions Janusz A. Starzyk, and Yue Li School of Electrical Engineering and Computer Science Ohio University,

15

Statistical Model of R-netsStatistical Model of R-nets

• The expected value of the number of different primary neurons reaching to (or reached from) a secondary neuron is

• The expected value of the number of secondary neurons reached from (or reaching to) a primary neuron is

• The number of primary neurons linked to a given primary neuron is

s

k

sp kP

Pks

111*

p

k

ps kS

Skp

111*

ps

kk

ci kkP

PPps

111*

Page 16: Neural Network with Memory and Cognitive Functions Janusz A. Starzyk, and Yue Li School of Electrical Engineering and Computer Science Ohio University,

16

Spurious neurons Spurious neurons

Spurious neuron • Is not a part of the original pattern

but is activated during the recall process.

• The probability that a potentially spurious neuron cj, is inhibited depends on the probability of an inhibitory link from an active primary neuron.

• Alternatively, a neuron may be spurious if it is not linked to any activated neurons at all.

spurious cjactive primary ci spurious

activated secondary b

trained projection

trained projection

pattern

Page 17: Neural Network with Memory and Cognitive Functions Janusz A. Starzyk, and Yue Li School of Electrical Engineering and Computer Science Ohio University,

17

Spurious neuronsSpurious neurons

A neuron is spurious if it satisfies the following conditions: a) it has no projection from Swa

b) all its projections from Ssa are trained.

Ssa is the strongly activated set of secondary neurons and

Swa is the weakly activated set of secondary neurons

C C-Cr r

S Ssa wa

Spurious node z

trained projections

Page 18: Neural Network with Memory and Cognitive Functions Janusz A. Starzyk, and Yue Li School of Electrical Engineering and Computer Science Ohio University,

18

Eliminating spurious neuronsEliminating spurious neurons

• The probability that a secondary node y belongs to Sa is

• The probability that a node in Sa is strongly activated is approximately

• The probability that a potentially spurious node, z, is not linked to any node in Swa is

1 1spk

rasr

Cp

P

( 1) 1

11 (1 ) sp srk P

ssa tP P

(1 (1 )) psk

nwa asr ssaP P P

Page 19: Neural Network with Memory and Cognitive Functions Janusz A. Starzyk, and Yue Li School of Electrical Engineering and Computer Science Ohio University,

19

Eliminating spurious neuronsEliminating spurious neurons

• The probability that a primary node z not linked to Swa is connected to a node in Ssa is

• The probability that z is a spurious node is

• The probability that the recall set has no more than Smax spurious neurons is max

_0

(1 )S

i P C iNO spurious z z

i

P CP P P

i

10

(1 )ps

ps

kk kps k k

z nwa pssa pssa tk

kP P P P P

k

1 (1 )asr ssa

pssaasr ssa

P PP

P P

Page 20: Neural Network with Memory and Cognitive Functions Janusz A. Starzyk, and Yue Li School of Electrical Engineering and Computer Science Ohio University,

20

Missing neuronsMissing neurons

• A missing neuron is a neuron from C-Cr which is suppressed by an inhibitory projection to an activated primary neuron.

• Lemma

Each missing neuron is suppressed by an inhibitory link to a spurious neuron connected through a secondary node w, where w is different from all nodes in activated set Sa.

C C-Cr

r

S Ssa wa

Spurious node z

m

w

untrained link

Page 21: Neural Network with Memory and Cognitive Functions Janusz A. Starzyk, and Yue Li School of Electrical Engineering and Computer Science Ohio University,

21

Eliminating missing neuronsEliminating missing neurons

• The probability that a given primary node will be missing due to a single spurious node can be estimated to be less than

• The probability that a single primary neuron is missing is

• The probability that the recall set has no more than Smax missing neurons is

sin _ 1 2

1(1 ) (1 )mis g one t t ps sp asr zP P P k k P P

P

sin sin _1 (1 )P Cmis g mis g oneP P

max

_ sin sin sin0

(1 )S

i C iNO mis g mis g mis g

i

CP P P

i

Page 22: Neural Network with Memory and Cognitive Functions Janusz A. Starzyk, and Yue Li School of Electrical Engineering and Computer Science Ohio University,

22

Results of the Statistical ModelResults of the Statistical Model

Storage capacities of R-nets with 4000 and 64,000 primary neurons

The statistical model is in a good agreement with simulated R-nets and can be ap

plied to estimate the computational performance of very large R-nets.

Page 23: Neural Network with Memory and Cognitive Functions Janusz A. Starzyk, and Yue Li School of Electrical Engineering and Computer Science Ohio University,

23

Results of the Statistical ModelResults of the Statistical Model

Storage capacities of R-nets for 100 and 150 neurons patterns

Page 24: Neural Network with Memory and Cognitive Functions Janusz A. Starzyk, and Yue Li School of Electrical Engineering and Computer Science Ohio University,

24

Results of the Statistical ModelResults of the Statistical Model

• Storage capacity of R-nets with 106 up to 107 primary neurons

• 105 to 2x106 secondary neurons

• 20 to 160 neuron patterns

• Up to 1000 projections from each primary neuron to secondary neurons with

• Up to 5000 projections from each secondary neuron back to primary neurons.

Page 25: Neural Network with Memory and Cognitive Functions Janusz A. Starzyk, and Yue Li School of Electrical Engineering and Computer Science Ohio University,

25

Simulations of R-netsSimulations of R-nets

0 20 40 60 80 100 1200

5

10

15

20

25

Set size

Set

s st

ored

storage capacity with recall size 75% ,50% and 30%

storage capacity for 1 cycle of recall of an R-net with 4000 primary neurons

Page 26: Neural Network with Memory and Cognitive Functions Janusz A. Starzyk, and Yue Li School of Electrical Engineering and Computer Science Ohio University,

26

Iterative operation of R-netsIterative operation of R-nets

Quick suppression of the spurious neurons

Spurious neurons suppressed after 14 cycles

Page 27: Neural Network with Memory and Cognitive Functions Janusz A. Starzyk, and Yue Li School of Electrical Engineering and Computer Science Ohio University,

27

Iterative operation of R-netsIterative operation of R-nets

Oscillations in the recovery process

Spurious neurons destroy the recovery process

Page 28: Neural Network with Memory and Cognitive Functions Janusz A. Starzyk, and Yue Li School of Electrical Engineering and Computer Science Ohio University,

28

ConclusionConclusion

• When the training sets are small, errors are due to the firing of spurious neurons with few links to the recall sets

• When the training sets are large, errors are chiefly due to over-training

• Storage capacity grows faster than the number of primary neurons with the slope of growth close to 10/7 on the logarithmic scale

• R-net with 106 primary neurons can store more than 130,000 100-neuron sets

• When R-net size reaches 109 primary neurons

(with average number of projections 104 - similar to interconnection density of human brain), the network can store over 109 patterns of 150 neurons each