ben goertzel monash 2011

Ben GoertzelCEO, Novamente LLC and Biomind LLC

CTO, Genescient CorpCo-founder, OpenCog ProjectVice Chairman, Humanity+

Adjunct Research Professor, Xiamen University, ChinaAdvisor, Singularity University and Singularity Institute

OpenCog:An Open Source Software Framework

&A Design & Vision for Advanced AGI

Aspects of AGI

• Philosophy of Mind

• Conception of General Intelligence

• Cognitive Architecture

• Software Architecture

• Environment & Tasks

• Developmental Roadmap

Philosophy of Mind

a mind is an evolving, autopoietic, self-referring set of patterns, associated with a system that’s interpreted as goal-achieving -- including patterns in the system and the world (and emergent

therebetween), and patterns regarding goal-achievement

2006 2010 2012 (?)

Conception of General Intelligence

Artificial General Intelligence (AGI)

Humans are, in a certain sense, general-purpose rather than narrowly specialized intelligences…

General intelligence may be loosely conceived as

“The ability of a system to achieve a variety of complex goals in a variety of complex environments using limited computational resources -- including goals and environments that were not anticipated at the time the system was created.”

memory

prediction

perception

action

goals

a mind uses

and

to do

of what

will achieve its

core operating principle of a general intelligence

Legg and Hutterʼs Definition of General Intelligence

Universal intelligence is a weighted average over all environments, of the intelligence of the agent in that environment

The weights assigned to environments are determined by the universal distribution, in which weight of an environment is (exponentially) inversely proportional to the length of the shortest program that computes it (on an assumed reference computer)

Goertzelʼs Definition of General Intelligence

The "pragmatic general intelligence" of an agent is the expected degree to which will achieve the goals specified in a certain goal distribution, in the environments specified in a certain environment distribution

Goertzelʼs Definition of Efficient General Intelligence

The "efficient pragmatic general intelligence" is the expected value over (goal, environment) pairs, of: the expected degree to which the agent will achieve the goal in the environment, normalized by the computational expense it will incur in doing so.

In real life, AGI is about making agents with high efficient pragmatic general intelligence relative to relevant classes of goals and environments

Defining the Intelligence of Real-World Agents

• Real-world intelligent systems do not always have explicit rewards or goals

• However, they can be locally approximated by goal-oriented systems, during each interval of time, and their intelligence can be estimated via the intelligence of these approximants

Defining the “Generality” of an Agentʼs Intelligence

The extent to which an agentʼs intelligence is general may be defined as the entropy of its environment-specific intelligence, evaluated according to a given prior distribution over environments

In this approach the generality of a systemʼs intelligence is orthogonal to its degree of intelligence

General intelligence in regard to the environments & goals for which humans evolved, and in which humans now operate, involves certain key competencies

Cognitive Architecture

Should we implement AGI via...

-- A single core complex cognitive process, supported by others as needed

-- Multiple simple processes interacting together, emergently yielding intelligence... ?

-- A number of complex cognitive processes, interacting together in a specific way? Glued together perhaps

using probabilistic & economic semantics?

The human brain seems involve a number of complex processes (modelable as algorithmic), interacting together in specified ways...

Which is critical?

-- The overall cognitive architecture?

-- The power of the learning algorithms?

-- Both? and they have to synergize well?

\

Which is critical?

-- The particulars of the system’s explicitly implemented processes and representations?

-- The structures and dynamics that emerge via complex self-organization, as the system grows and

learns?

-- Both? and they have to synergize well?

\

MotorSensory

Declarative Episodic AttentionalIntentional Procedural

OpenPsi

DeSTIN(vision / audition)

RewardHierarchy

Motor Control

Hierarchy

robot proxy

game proxy

memory

PLN (inference)

Pattern mining

conceptblending

dimensional embedding

MOSES Hillclimbing Internal simulation ECAN

(economic attention allocation

Language processing

DeclarativeProbabilistic Logic

Networks,concept blending,

language comprehension &

generation

Episodicinternal world

simulation engine

Attentional/Intentional

economic attention networks, adaptive

goal hierarchy

Sensoryhierarchy of memory/

processing units

ProceduralMOSES

(probabilistic evolutionary

learning),hillclimbing

cognitive synergy in OpenCog

In OpenCog, multiple cognitive processes act concurrently on the

same knowledge store

Explicit knowledge representation:

Nodes and links (collectively “Atoms”) that explicitly encode individual pieces of knowledge

Implicit knowledge representation:

Knowledge that is encoded in the coordinated structure or activity of a large set of nodes and links

Knowledge Representation in OpenCog

Nodes are typed and may •symbolize entities in the external world embody simple executable processes•symbolize abstract concepts•serve as components in relationship-webs signifying complex concepts or procedures.

Links are typed and may•be binary, unary or n-ary•point to nodes or links;•embody various types of relationships between concepts,percepts or actions.•The network of links is a web of relationships.

The two types of OpenCog Atoms: Nodes and Links

Atoms come with TruthValue and AttentionValue objects

TruthValue objects come in several forms, e.g.

Single probability values

SimpleTruthValues, consisting of (probability, count) pairs

IndefiniteTruthValues, consisting of tuples (Lower bound, Upper bound, confidence)

DistributionalTruthValues, each encompassing a first or second order probability distribution

AttentionValue objects contain information telling how much processor time and memory an Atom should get. Most simply an AttentionValue object contains

ShortTermImportance (STI) value telling how much processor time the Atom should get in the near future

LongTermImportance (LTI) value telling how important it is to retain the Atom in memory

VLTI bit telling whether, if the Atom is removed from RAM, it should be kept on disk

There are multiple types of Nodes

Each type has its own semantics

Informally, we may divide the node types into multiple “varieties”

The precise collection of Atom types has changed over the years as we have developed the OpenCog system

Someday, a future version may learn its own Atom types

There are multiple types of LinksEach type has its own semantics

Informally, we may divide the link types into multiple “varieties”

Member Ben_Goertzel Goertzel_Family

Member Anchovy_Goertzel Goertzel_Family <.8>

("Anchovy" is a rabbit, so her membership degree in the Goertzel family is less than 1)

AsymmetricHebbianLink Ben_Goertzel insane <.5>

(this means that when "Ben Goertzel" is mentioned, "insane" is often also mentioned)

SymmetricHebbianLink dog pet

SymmetricHebbianLink ConceptNode: dog ConceptNode: pet

ContextLink USA SymmetricHebbianLink dog pet <.8>

ContextLink Korea SymmetricHebbianLink dog pet <.5>

ExecutionLink + ListLink 2 3 5

ExecutionLink SchemaNode: + ListLink NumberNode: 2 NumberNode: 3 NumberNode: 5

ExecutionLink kick ball_44

ExecutionLink SchemaNode: kick SemeNode: ball_44

ExecutionLink give ListLink (ChenShuo, ball_44)

InheritanceLink cat animal

InheritanceLink animal cat <.01>

InheritanceLink animal cat <.01,.95>

SubsetLink parrot human <0>

IntensionalInheritanceLink parrot human <.6>

IntensionalInheritanceLink frog human <.03>

EvaluationLink <.8> PredicateNode: is_silly SemeNode: Ben_Goertzel

ForAllLink VariableNode: $X, $Y ImplicationLink <.7> AndLink InheritanceLink $X man InheritanceLink $Y woman EvaluationLink love ListLink ($X, $Y) EvaluationLink overestimate <.9> ListLink $X EvaluationLink beauty $X

Ben Goertzel is silly

When a man loves a woman, he greatly overestimates her beauty

MindAgents are objects that act on the Atomspace: modifying, adding and/or removing Atoms

MindAgents’ activities are scheduled by a Scheduler: the simplest Scheduler would just cycle through all available MindAgents

OpenCog involves a carefully constructed combination of MindAgents, intended to reinforce rather than confuse each other!

MOSES Probabilistic Evolutionary Learning

Combines the power of two leading AI paradigms: evolutionary and probabilistic

learning

Extremely broad applicability. Successful track record in bioinformatics, text and data mining,

and virtual agent control.

Moshe Looks 2006 PhD thesis: metacog.org

Probabilistic Logic NetworksA highly general, practical integration of

probability theory and symbolic logic.

Extremely broad applicability. Successful track record in bio text mining, virtual agent control.

Based on mathematics described in Probabilistic Logic Networks, published by Springer in 2008

Two Key Algorithms for Procedural and Declarative Knowledge Creation

Concept Blending

Map Encapsulation

Economic Attention Allocation

Each node or link in the AtomSpace is tagged with a probabilistic truth value, and also with an “attention value”, containing Short-Term Importance and Long-Term Importance components.An artificial-economics-based process is used to update these attention values dynamically -- a complex, adaptive nonlinear process.

Psi Architecture Overviewfigure from “Principles of Synthetic Intelligence” by Joscha Bach

Symbol Grounding in Psifigure from “Principles of Synthetic Intelligence” by Joscha Bach

Joscha Bach’s MicroPsi Architecturefigure from “Principles of Synthetic Intelligence” by Joscha Bach

Basic net-entity from which MicroPsi’s node types,comprising its dynamic

knowledge representation,are composed

Psi OpenCog

Memory AtomSpace (weighted semantic hypergraph), Procedure repository, space/time index servers, etc.

Demands Implemented as GroundedPredicateNodes

Urges Called “Ubergoals”, these are also GPNs, with truth values calculated in terms of corresponding Demands

Urgency For goal Atoms, the ShortTermImportance value indicates urgency

Pleasure A GPN whose internal TV evaluation function compares actual to expected levels of Ubergoal satisfaction

Goals Atoms (both Ubergoals and learned subgoals) in the system’s GoalPool

Motive Selection Carried out as a function of Economic Attention Allocation, which dispenses STI funds to goals

Action Selection Similar to in Psi, OpenPsi selects composite procedures to execute based on its inferences of what may best achieve its goals

Planning Arises as a consequence of multiple cognitive processes, including PLN, MOSES and ECAN

Modulators (e.g. activation, resolution level, certainty,

selection threshold)

System parameters, representable as GPNs, with TVs estimated by appropriate heuristic formulae

Other approaches, e.g. Itamar Arel’s DeSTIN, follow similar principles with different details -- a general term I’ve used is “compositional spatiotemporal

deep learning networks”. These may be hybridized with OpenCog.

adviceinput

below

belowbelow

between

aligned on axis perpendicular to

mouth-nose-eyes axis...

visual patterns (e.g. DeSTIN centroids) corresponding to

the semantic perception node linked to “human

eye”

concept node

phrase node

“human eye”

referencelink

PERCEPTUAL CSDLN

SEMANTIC-PERCEPTUAL CSDLN

COGNITIVESEMANTICNETWORK

referencelink

referencelink

semantic HTM nodes/linksformed by pattern mining

perceptual HTM

semantic HTM provides probabilistic biasing to

perceptual HTM

aligned on axis perpendicular to

eyes axis

CSDLN = “Compositional Spatiotemporal Deep Learning

Network, e.g. HTM, DeSTIN,...

grasp object

The motoric hierarchy node corresponding to a particular set of servomotors, might for instance contain

clusters of paths through configuration space that the servomotors have historically followed.

MOTORIC CSDLN

SEMANTIC- MOTORIC CSDLN

rotate arm toward object

raise or lower arm toward object

raise or lower arm toward default position

after

after

before/after/simultaneous

get object

adviceinput


“get”

phrase nodeconcept node

referencelink

referencelink

referencelink

motor patterns (e.g. centroids) corresponding to the semantic motoric

node “raise or lower arm toward object”

SEMANTIC- MOTORIC CSDLNSEMANTIC-PERCEPTUAL CSDLN

SEMANTIC GOAL CSDLN


possess food

eat food

maintainappropriate

fullness

get object(e.g. food)

see food & get food --> possess food

see food

see meat see sandwich

... ...

“get”Hungry & eat food--> maintain appropriate fullness

How --->

Helps |

\|/

Map formation Goal system Simulation Sensorimotor

pattern recognition

Uncertain inference Creates new concepts and relationships, enabling briefer useful inference trails

Goal refinement enables more careful goal-based inference pruning

- Simulations provide a method of testing speculative inferential conclusions

- Simulations suggest hypotheses to be explored via inference

Creates new concepts and relationships, enabling briefer useful inference trails

Supervised procedure learning

Creates new procedures to be used as modules in candidate procedures

Goal refinement allows more precise definition of fitness functions, making procedure learningʼs job easier

Simulation provides a method of “fitness estimation” allowing inexpensive testing of candidate procedures

Extraction of sensorimotor patterns allows creation of abstracted fitness functions for (inferentially and simulatively) evaluating procedures guiding real-world actions

Attention allocation Creates new concepts grouping “attentionally related” memory items, enabling AA to find subtler attentional patterns involving these nodes

Goal refinement allows more accurately goal-driven allocation of attention

Simulation provides data for attention allocation -- allowing attentional information to be extracted from co-occurences observed in simulation

Creates concepts grouping “attentionally related” memory items, enabling AA to find subtler attentional patterns involving these nodes

Concept creation Creates new concepts to be fed into other concept creation mechanisms

Goal refinement provides more precise definition of criteria via which new concepts are created

Utility of concepts may be assessed via creating simulated entities embodying the new concepts and seeing what they lead to in simulation

Creates new concepts to be fed into other concept creation mechanisms

How --->

Helps |

\|/

Uncertain inference Supervised procedure learning

Attention allocation Concept creation

Uncertain inference NA When inference gets stuck in an inference trail, it can ask procedure learning to learn new patterns regarding concepts in the inference trail (if there is adequate data regarding the concepts)

Importance levels allow pruning of inference trees

Provides new concepts, allowing briefer useful inference trails

Supervised procedure learning

Inference can be used to allow prior experience to guide each instance of procedure learning.

NA Importance levels may be used to bias choices made in the course of procedure learning

(e.g. in OCP, in the fitness evaluation and representation-building phases of MOSES)

Provides new concepts, allowing compacter programs using new concepts in various roles

Attention allocation Enables inference of new HebbianLinks and HebbianPredicates from existing ones

Procedure learning can recognize patterns in historical system activity, which are then used to build concepts and relationships guidng attention allocation

NA Combination of concepts formed via map formation, may lead to new concepts that even better direct attention

Concept creation Allows inferential assessment of the value of new concepts

Procedure learning can be used to search for high-quality blends of existing concepts (using e.g. inferential and attentional knowledge in the fitness functions)

Allows assessment of the value of new concepts based on historical attentional knowledge

NA

How --->

Helps |

\|/

Uncertain inference Supervised procedure learning

Attention allocation Concept creation

Map formation Speculative inference can help map formation guess which maps to hunt for

Procedure learning can be used to search for maps that are more complex than mere “co-occurrence”

Attention allocation provides the raw data for map formation

No significant direct synergy

Goal system Inference can carry out goal refinement


Flow of importance among subgoals determines which subgoals get used, versus being forgotten

Concept creation can be used to provide raw data for goal refinement (e.g. a new subgoal that blends two others)

Simulation In order to provide data for setting up simulations, inference will often be needed


Attention allocation tells which portions of a simulation need to be run in more detail


Sensorimotor

pattern recognition

Speculative inference helps fill in gaps in sensory data

Procedure learning can be used to find subtle patterns in sensorimotor data

Attention allocation guides pattern recognition via indicating which sensorimotor stimuli and patterns tend to be associatively linked

New concepts may be created that then are found to serve as significant patterns in sensorimotor data

How --->

Helps |

\|/

Map formation Goal system Simulation Sensorimotor

pattern

recognition

Map formation NA Map formation may focus on finding maps related to subgoals, and good subgoal refinement helps here



Goal system Concepts formed from maps may be useful raw material for forming subgoals

NA No significant direct synergy


Simulation No significant direct synergy


NA Presence of recognized sensorimotor patterns may be used to judge whether a simulation is sufficiently accurate

Sensorimotor

pattern

recognition

Concepts formed from maps may usefully guide sensorimotor pattern search

Directing pattern search toward patterns pertinent to subgoals, may make the task far easier

Patterns recognized in simulations may then be checked for presence in real sensorimotor data

NA

Software Architecture

Currently the OpenCog “core (AtomSpace, scheduler, comms,

saving to disk, etc.) is C++ for Unix, with an STL-based API

MindAgents are currently

coded in C++ or Python.

Some MindAgents also invoke

external Java processes

Environment & Tasks

2009 AGI Roadmap Workshop

How Important Is Embodiment?

Some AI theorists believe that robotic embodiment is necessary for the achievement of powerful AGI

Others believe embodiment is entirely unnecessary

We believe embodiment is extremely convenient for AGI though perhaps not strictly necessary; and that virtual-world embodiment is an important, pragmatic and scalable approach to pursue alongside physical-robot embodiment

Current virtual world platforms have some fairly severe limitations, which fortunately can be remedied with effort

Object-object interactions are oversimplified, making tool use difficult

Agent control relies on animations and other simplified mechanisms, rather than having virtual servomotors associated with each joint of an agent’s skeleton

Partial solution: Integration of a robot simulator with a virtual world engine

Player / Gazebo: 3D robot control + simulation framework

OpenSim: open-source virtual world

It seems feasible to replace OpenSim’s physics engine with appropriate components of Player/Gazebo, and make coordinated OpenSim client modifications

+

Current Virtual Worlds lack fluids, powders, pastes, fabrics … they donʼt completely implement

“naïve physics”

One likely solution: bead physics

Spherical beads with specially designed adhesion properties can emulate fluids, fabrics, pastes, strings, rubber bands, etc.

Bead physics can be added to virtual world physics engines

Possible solution: build robots from “macrocells” -- flexible ball-like units that can stretch into different shapes, sense the environment, and pass power and

information to each other

Current robots are made mainly of inert parts -- very few of their parts are sensors or actuators...

This shows OpenCog oscillating between the robot's "home" and batteries, as a result of its quests to fulfill integrity and energy demands respectively. After a few movements back and forth, it also shows the Psi monitor updating graphs of the Psi variables (note the updates are synchronized between graphs).

This video shows OpenCog spelled out in blocks, as a 3D path finding mission. It also shows the demand for finding a battery to restore energy.

Subsequent to finding the battery, the avatar asks the player for a battery. The player then pushes a button to make one appear. A video transition happens where some footage was cut out (involving more observations of the button generating food), and the video then shows the OpenCog avatar using the button to create food for itself.

Developmental Roadmap

Piagetan Stages of Development

Piaget Meets Uncertain Inference

• 2011-2012: A Proto-AGI Virtual Agent

• 2013-2014: A Complete, Integrated Proto-AGI Mind (Piagetan concrete)

• 2015-2016: Advanced Learning and Reasoning (Piagetan formal)

• 2017-2018: AGI Experts

• 2019-2021: Full-On Human Level AGI

• 2021-2023: Advanced Self-Improvement (Piagetan reflexive)

• 6 month from start (now): Unity-based minecraft-like world initially built, proxy to Unity built, 3D pathfinding, frequent subgraph miner designed/built, PLN and “embodiment” (body-control) systems debugged/refactored, PLN based planner built, OpenPsi motivation/emotion system implemented

• 12 month (end of 2011): New artwork/animations integrated into world, further behaviors/interactions for objects scripted/tested, PLN further refactored and simple temporal/spatial reasoning integrated, reinforcement/imitation learning integrated (and interface btw this learning and OpenPsi created), dialogue system refactored/improved, appropriate emotional responses demostrated. Reliable live demo!

• 18 month (mid-2012): Attention allocation integrated, complex planning demonstrated, learning & reasoning integrated, linguistic question answering and command demonstrated. Exciting live demo...

• 24 month (end of 2012): Software toolkit built out to enable tractable integration into games. Demo improvements. OpenCog 1.0 release. Intelligence improvements

OpenCog Hong Kong Project

• 2010: Simplistic integration of OpenCog with Nao robot for navigation, command-following; implementation of English language generation for OpenCog

• 2011: Re-implementation of Itamar Arel’s team’s DeSTIN architecture in CUDA for GPU computing: initial version completed June 2011, refinements underway

• 2012-13: Integration of DeSTIN with OpenCog using intermediate “semantic compositional spatiotemporal deep learning network”; integration of robotic movement control system with OpenCog

• 2013-2014: OpenCog-powered intelligent robotics

Xiamen University BLISS Lab

Interim Applications

biomedicalinformatics

financial prediction /

analytics

video games / virtual worlds

robot toys

service robots

information retrieval

etc. !!!

The natural synergy between advanced AI and gaming/virtual worlds has been avidly discussed for at least a decade, and is now finally becoming a practical reality

For a readable journalistic treatment of the application of machine learning to study longevity in

fruit flies, see

“AIs, Superflies and the Path to Immortality”

in H+ Magazine, hplusmagazine.com

Ensemble based machine learning does well here, but appropriate OpenCog integration could do much better by allowing a vast variety of available bio

datasets to be brought to bear on the analysis of any one dataset.

http://hplusmagazine.com

http://hplusmagazine.com

ben goertzel monash 2011

Science