19
© NOKIA cm architecture.PPT / 10.8.2003 / PHa page: 1 An Implementable Architecture for Conscious Machines Dr. Pentti O A Haikonen, Principal Scientist, Cognitive technology Nokia Research Center

© NOKIAcm architecture.PPT / 10.8.2003 / PHa page: 1 An Implementable Architecture for Conscious Machines Dr. Pentti O A Haikonen, Principal Scientist,

Embed Size (px)

Citation preview

Page 1: © NOKIAcm architecture.PPT / 10.8.2003 / PHa page: 1 An Implementable Architecture for Conscious Machines Dr. Pentti O A Haikonen, Principal Scientist,

© NOKIA cm architecture.PPT / 10.8.2003 / PHa page: 1

An Implementable Architecture for Conscious Machines

An Implementable Architecture for Conscious Machines

Dr. Pentti O A Haikonen, Principal Scientist, Cognitive technology

Nokia Research Center

Page 2: © NOKIAcm architecture.PPT / 10.8.2003 / PHa page: 1 An Implementable Architecture for Conscious Machines Dr. Pentti O A Haikonen, Principal Scientist,

© NOKIA cm architecture.PPT / 10.8.2003 / PHa page: 2

Introduction

The aim of my research is to develop a conscious machine; not one with a

“flea’s consciousness” but a robot brain with the cognitive abilities and

hallmarks of the human conscious mind.

This machine would have to posses:

-Awareness of its environment, time and place

-Self-consciousness; awareness of its existence as an independent entity,

awareness of its mental content like inner speech and inner imagery and

the recognition of these as its own

-Apparent subjective immateriality of the mental content

-Cognitive functions that parallel those of the human brain; symbol

processing in the human sense, natural language

-Emotions, emotional significance and judgement

Page 3: © NOKIAcm architecture.PPT / 10.8.2003 / PHa page: 1 An Implementable Architecture for Conscious Machines Dr. Pentti O A Haikonen, Principal Scientist,

© NOKIA cm architecture.PPT / 10.8.2003 / PHa page: 3

Environment Perception process Internal process

percepts of:-environment-inner states

objects

actions

situationmatch/mismatch/novelty detection

tasks, goals, needsmemories

learned routines

predictionreasoningplanned actionjudgement

ReactionsActions

etc.

experience

relationships

emotionsemotionalevaluation

The General Model of Cognition

“mental loop”

“external loop”

Page 4: © NOKIAcm architecture.PPT / 10.8.2003 / PHa page: 1 An Implementable Architecture for Conscious Machines Dr. Pentti O A Haikonen, Principal Scientist,

© NOKIA cm architecture.PPT / 10.8.2003 / PHa page: 4

-Processing with meaning and significance

-Distributed neural signal array representation

-Associative soft symbol processing

-System reactions; pain, pleasure, good, bad, emotions

-Match/mismatch/novelty detection

-Distributed attention controlled by significance

-Modulation domain operation

Basic Specifications I

Page 5: © NOKIAcm architecture.PPT / 10.8.2003 / PHa page: 1 An Implementable Architecture for Conscious Machines Dr. Pentti O A Haikonen, Principal Scientist,

© NOKIA cm architecture.PPT / 10.8.2003 / PHa page: 5

-Perception process, incl. attention, priming and prediction

-Learning, also learning by imitation, “mirror neurons”

-Inner speech, natural language

-Inner imagery

-Introspection, judgement of own thoughts

-Imagination, planning (involves the imagery of “self” executing

imagined actions)

-True “immaterial” consciousness; self-awareness

Basic Specifications II

Page 6: © NOKIAcm architecture.PPT / 10.8.2003 / PHa page: 1 An Implementable Architecture for Conscious Machines Dr. Pentti O A Haikonen, Principal Scientist,

© NOKIA cm architecture.PPT / 10.8.2003 / PHa page: 6

environment

sensors & pre-processing

rawdistributed signals

innerprocesses

feedback

perception process

percepts

responses

The Outline of a Conscious Machine

Page 7: © NOKIAcm architecture.PPT / 10.8.2003 / PHa page: 1 An Implementable Architecture for Conscious Machines Dr. Pentti O A Haikonen, Principal Scientist,

© NOKIA cm architecture.PPT / 10.8.2003 / PHa page: 7

Neuron groups for Inner Processes

“Neuron group”Main input signal array

Output signal array

Associative signal array inputs

-Can associatively connect large arrays of neural signals to each other

-Can associatively connect neural signal array sequences to each other

-Can associatively evoke arrays of neural signals and neural signal array

sequences even by limited length and incomplete cues

-Sensitive to signal intensity; significance control

-Can determine match/mismatch/novelty conditions between main input

signal array and associative signal arrays

-Finds “best fit” by Winner-Takes-All threshold operation

Page 8: © NOKIAcm architecture.PPT / 10.8.2003 / PHa page: 1 An Implementable Architecture for Conscious Machines Dr. Pentti O A Haikonen, Principal Scientist,

© NOKIA cm architecture.PPT / 10.8.2003 / PHa page: 8

The General Architecture

perceptpreprocesssensor

preprocesssensor

preprocesssensor

preprocesssensor

Body position

Touch subsystem

Auditory subsystem

Visual subsystem

system reactions

System

perception process

perception process

perception process

perception process

perception process

neuron group

neuron group

neuron group

neuron group

neuron groupsmell, taste, pain, etc.

motor actions threshold

threshold

percept

percept

percept

percept

Page 9: © NOKIAcm architecture.PPT / 10.8.2003 / PHa page: 1 An Implementable Architecture for Conscious Machines Dr. Pentti O A Haikonen, Principal Scientist,

© NOKIA cm architecture.PPT / 10.8.2003 / PHa page: 9

The Visual Subsystemperception process

perception process

perception process

perception process

neuron group

neuron group

neuron group

neuron group

shape

size

color

perception process

neuron group

motion

gaze direction control

visualsensors

position

shape

size

color

motion

position

percept

percept

percept

percept

percept

Page 10: © NOKIAcm architecture.PPT / 10.8.2003 / PHa page: 1 An Implementable Architecture for Conscious Machines Dr. Pentti O A Haikonen, Principal Scientist,

© NOKIA cm architecture.PPT / 10.8.2003 / PHa page: 10

The General Architecture II

-Each modality works on its own and produces streams of percepts about

environment and internal states.

-Modalities are associatively cross-connected, therefore the activity of one

modality may be reflected in the other modalities; percepts may be named

and labeled, names may evoke corresponding percepts…the activity of one

modality may be memorized and reported in terms of other modalities, etc.

-Attention determines which percepts are accepted for further action.

Attention is controlled by signal intensity and thresholds, these are

controlled by e.g. emotional significance.

-Pain and pleasure are system reactions that affect attention.

Page 11: © NOKIAcm architecture.PPT / 10.8.2003 / PHa page: 1 An Implementable Architecture for Conscious Machines Dr. Pentti O A Haikonen, Principal Scientist,

© NOKIA cm architecture.PPT / 10.8.2003 / PHa page: 11

Why Would This Machine be Conscious?

We are conscious when we are able to report to ourselves what we are

experiencing and are able to make memories of that.

This machine is able to produce these reports; it can make a verbal note e.g.

about what is being seen and it can also make an episodic memory of that.

The machine will accumulate a personal history.

A self-conscious being must be able to perceive the difference between the

percepts caused by external entities and percepts of its own inner activity.

There are several ways to achieve this within this architecture.

The machine will, in principle, be able to report having a flow of inner speech

and inner imagery and will claim the ownership of these.

The “mental activity” takes place in modulation domain, hence the mental

content appears as immaterial.

Page 12: © NOKIAcm architecture.PPT / 10.8.2003 / PHa page: 1 An Implementable Architecture for Conscious Machines Dr. Pentti O A Haikonen, Principal Scientist,

© NOKIA cm architecture.PPT / 10.8.2003 / PHa page: 12

The purpose of the work is to develop associative neuron group microchips that

-Are suitable for cognitive and conscious neural system architectures

-Can learn; associatively connect very large arrays of neural signals to each other

-Can associatively connect neural signal array sequences to each other

-Can associatively evoke arrays of neural signals and neural signal array sequences by limited length and incomplete cues (soft symbol processing)

-Can accommodate importance

-Can determine match/mismatch/novelty conditions between neural signal arrays

NRC associative neuron group integrated circuit development work

Page 13: © NOKIAcm architecture.PPT / 10.8.2003 / PHa page: 1 An Implementable Architecture for Conscious Machines Dr. Pentti O A Haikonen, Principal Scientist,

© NOKIA cm architecture.PPT / 10.8.2003 / PHa page: 13

Experiments with the neuron microchip ver. 2001

This test system could learn associatively short melodies and mimic them

when few notes were played as a cue. The system learned the pitches of the

notes as well as their duration; also the duration of any silent(!) interval.

Interval detector

Microphone + one octave filter bank

Audio oscillator bank

Associative neuron groups and perception-response loops

Page 14: © NOKIAcm architecture.PPT / 10.8.2003 / PHa page: 1 An Implementable Architecture for Conscious Machines Dr. Pentti O A Haikonen, Principal Scientist,

© NOKIA cm architecture.PPT / 10.8.2003 / PHa page: 14

Associative neuron group integrated circuit v. 2002

An Experimental Integrated Circuit for Conscious Machines

-24 neurons, 24 x 24 synapses in 8 individual groups (total 4608)

- Fully parallel internal operation, serial external communication

-Contains the main features of a future practical large scale microchip

s

soP/S

M

MM

N

S/P S/P S/P

threshold threshold threshold

TH

threshold

n x

neuron body

a1 a2 am

n x n synapses n x n synapses n x n synapses

enab1 enab2 enabm

lc

ck1 ck2

priority

enabWTAs_ctr

reset

S/P

M/MM/N circuit

Page 15: © NOKIAcm architecture.PPT / 10.8.2003 / PHa page: 1 An Implementable Architecture for Conscious Machines Dr. Pentti O A Haikonen, Principal Scientist,

© NOKIA cm architecture.PPT / 10.8.2003 / PHa page: 15

Associative neuron group integrated circuit ver. 2002

Page 16: © NOKIAcm architecture.PPT / 10.8.2003 / PHa page: 1 An Implementable Architecture for Conscious Machines Dr. Pentti O A Haikonen, Principal Scientist,

© NOKIA cm architecture.PPT / 10.8.2003 / PHa page: 16

-Original circuitry and specifications developed by the author at NRC

-Actual implementation on silicon executed by VTT Microelectronics

Neuron chip ver. 2002 installed in test circuit board (square chip in the middle)

Associative neuron group integrated circuit ver. 2002

Page 17: © NOKIAcm architecture.PPT / 10.8.2003 / PHa page: 1 An Implementable Architecture for Conscious Machines Dr. Pentti O A Haikonen, Principal Scientist,

© NOKIA cm architecture.PPT / 10.8.2003 / PHa page: 17

Conclusions

The author has developed an architecture for conscious machines.

More detailed description and discussion can be found in the book:

The Cognitive Approach to Conscious Machines, Imprint Academic 2003

Experimental integrated circuits are being developed for the actual

implementation.

More work will be needed in the development of suitable input/output

sensors and devices and especially in the sensory signal preprocessing

area.

Page 18: © NOKIAcm architecture.PPT / 10.8.2003 / PHa page: 1 An Implementable Architecture for Conscious Machines Dr. Pentti O A Haikonen, Principal Scientist,

© NOKIA cm architecture.PPT / 10.8.2003 / PHa page: 18

An Implementable Architecture for Conscious Machines

An Implementable Architecture for Conscious Machines

Dr. Pentti O A Haikonen, Principal Scientist, Cognitive Technology

Nokia Research Center

Page 19: © NOKIAcm architecture.PPT / 10.8.2003 / PHa page: 1 An Implementable Architecture for Conscious Machines Dr. Pentti O A Haikonen, Principal Scientist,

© NOKIA cm architecture.PPT / 10.8.2003 / PHa page: 19

-end-