Upload
others
View
19
Download
1
Embed Size (px)
Citation preview
The next wave of radical AI progress:
The transition from narrow AI to
artificial general intelligence (AGI)
“We have reviewed progress
in the field over the past six
years. Our conclusion is
near term AGI should be
taken as a serious
opportunity. – Ilya SutskeverOpenAI
Founder and Research Director
1970s
2000s - 2010s
THE FUTURE
The AI field shifts towards Narrow AI application due to the
apparent difficulty of realizing early AGI visions
Cheap powerful computing and abundant data fuel a
commercial Narrow AI Revolution, focused heavily on
neural net algorithms analyzing huge datasets
Limitations of neural net data analytics becomes clear,
attention begins shifting to neural-symbolic method and the
search for new ideas
Russian National AGI Initiative
AGI begins to garner serious attention in industry and
government
Steady Narrow AI research progress,
most funding and applications are military
Cogistry moving toward AGI via Narrow AGI and
Computational Creativity
AGI
Historical
Timeline
1980s - 90s
Google invests $500+ million in Deep Mind
Microsoft invests $1 billion in OpenAI
2014
2019
Narrow AISpecific to narrow domains
Levels of AI
Artificial Super IntelligenceBeyond human-level AGI
Artificial General IntelligenceAbility to learn new domains and create
imaginatively
Narrow AGIGeneral learning ability within a domain
Major Roadblocks Remain Between
Here and Global AGI
No CooperationInteroperation of AIs
is unsystematic and
awkward
Top-Down
ControlA few large
corporations and
governments
control the AI scene
Weak
GeneralizationNarrow AIs are poor at
dealing with novel
situations outside their
training contexts
From Narrow AI to AGI
via Narrow AGI
● The path from here to AGI will be carried out
largely via the creation of “Narrow AGI” systems
that serve practical functions in specific
application areas
● A Narrow AGI service-robotics system would
display near human-like general intelligence in
the home and office service robotics domain, but
not necessarily much expertise outside this
domain of knowledge
● Similarly one may have biomedical Narrow AGIs,
Smart City Controller Narrow AGIs, etc.
● Connecting together multiple Narrow AGIs is one
path to a full-scale AGI
● To make this connection work AGI systems will
require components with everyday human
commonsense knowledge and practical
creativity
We at SingularityNET believe we have the
critical ingredients for advancing from
narrow AI to narrow AGI
Historically, the AI field has fostered a variety of
paradigms.
In recent years deep neural networks have become
commercially dominant, due largely to their success at
visual and auditory analysis and generation tasks.
Today there is an increasing realization that deep
neural nets will not suffice for AGI, and that these
useful narrow tools will have to be combined with
different AI modalities to achieve generalization and
creativity.
Some recent “deep neural net” successes are actually
hybrid architectures:
● AlphaGo architecture from Google DeepMind
integrates two neural nets with one game tree
● ERNIE architecture for Natural Language
Processing question-answering from
Tsinghua University integrates knowledge
graphs into neural networks
LOGIC
SYSTEMS
EVOLUTIONARY
LEARNING
ARTIFICIAL
LIFE
Shift Toward Hybrid
Architecture
NEURAL
NETWORKS
OpenCog, founded 2008, is a hybrid neural-symbolic AGI
architecture that stores its learned knowledge in a hypergraph
knowledge base known as its Atomspace.
The largest global AGI project outside of
big tech companies
Integrates logic systems, neural
networks, and evolutionary learning in a
single architecture, with multiple AI
algorithms acting on the same dynamic
knowledge hypergraph known as the
Atomspace
OpenCog AGI Platform
Integrated into SingularityNET platform
Computational Creativity via Cogistry
● Algorithmic chemistry was first articulated
by Walter Fontana, and the basic concept is
to create a collection of machine
instructions that are able to symbiotically
interact with each other
● In algorithmic chemistry, Machine Code A
can do work on Machine Code B to
transform it into Machine Code C – and
these Machine Codes all exist together
within a ‘primordial algorithmic soup’,
where they can act on each other and
create new interesting and dynamic
structures
● Cogistry involves a ‘self-modifying
algorithmic soup’ existing in OpenCog’s
Atomspace that is enhanced by advanced
AI algorithms such as hypergraph pattern
mining and probabilistic reasoning
Adds a self-organizing
imagination and creativity
component to the OpenCog
architecture
Cogistry is an attempt to
create new
AI/AGI/Artificial-Life on
OpenCog, through the
concept of algorithmic
chemistry
CREATIVITY
ALGORITHMIC
CHEMISTRY
Blockchain/Distributed
Ledger + Off-chain
Messaging/Storage
Infrastructure Abstraction
Data
AI Algorithms
Data Processing
Tools
AI Solutions/Services
Software & Hardware
Applications
Market
AGI Token
AI nodes may join
together into
“federations” –
subnetworks of nodes
that habitually work
together to carry out
particular sorts of tasks
All nodes in the SingularityNET carry out diverse
cognitive and analytic operations, and exchange diverse
types of data
This is a “network
effect” from both
business and cognitive
perspectives
Video Analysis Node
Voice Samples in need
of Sentiment Analysis
Voice Analysis AI Node
Semantic Analysis of
Sentences
Sentences in need of
Semantic Analysis
Natural Language
Processing AI Node
Sentiment Labeled
Voice SamplesList of Objects in
Video
Video Clips in Need
of Object
Recognition
1
5TODALARITY: AI + Blockchain Startup
Accelerator•
Co-founded by industry pioneers
• SingularityNET, the leading decentralized AI platform
• Toda, the breakthrough scalable secure value
transaction platform•
Providing seed funding, wisdom and hands-on help to startups
• based anywhere on the planet
• leveraging SingularityNET, Toda, and/or other leading
technologies at the intersection of AI and blockchain•
Accelerating the growth of the decentralized AI ecosystem
Loving AI"robots that respond with
unconditional love"
Hanson Robotics humanoid robot “Sophia”
has been used experimentally as a
meditation coach —
and pilot psychological studies
are very promising!
Robot Preschool:
One Initial AGI Testing Ground
● One way to proceed toward AGI is via creating a Narrow
AGI system demonstrating human-like general
intelligence in a “robot preschool” context
● A wheeled humanoid robot with flexible hand-like
manipulators and high-quality video and audio sensors, in a
“preschool” type environment, is within the scope of
commodity technologies and can easily be built in a lab for
research use or mass-produced commercially
● SingularityNET founders led the Sophia robot software
team at Hanson Robotics, and bring this experience to
Cogistry
● This system will then be integrated with applied
commercial narrow AI systems built using OpenCog,
yielding Narrow AGIs for other domains
From Robot Preschool to
Commercial Narrow AGI
Robot school can be used to train commercial robotic products, such as
domestic robot helpers and construction workers
AGI PRESCHOOL TRAIN VERTICAL
APPLICATIONS
COMMERCIAL
NARROW AGI
In 2018-19, OpenCog and SingularityNET partnered with
Cisco to build a neural-symbolic reasoning system for
traffic analytics in smart cities.
The team built a system using OpenCog’s abductive
reasoning to analyze street scenes and re-identify objects
within a smart city across non-overlapping cameras.
This neural-symbolic reasoning system did not just
populate a symbolic knowledge base with labels produced
by DNNs in a bottom-up way. Instead, it performed
grounded reasoning, guided by both data and seed
ontology, over symbols connected to subsymbolic
data, assembling module networks on the fly.
Such tight integration of deep neural networks and
symbolic reasoning allowed OpenCog to detect
undesirable behavior, such as jaywalking and possible
collisions alerts, and identify anomalies by creatively
discovering surprising patterns.
Neural-Symbolic AGI: OpenCog Smart City Application
Supercentenarians (>105)
Data provided by James Clement and
supercentenarian.com.
27 whole genome sequences taken from
subjects 105 – 117 years old. 14 women & 13
men of European ancestry.
Control group (chosen by SingularityNET) –
503 European samples from 1000 Genomes.
GWAS shows no genome-wide significant
SNPs
Extreme longevity phenotype is associated with above
average genome-wide SNP density
BETTER HUMANS vs CONTROL The mean SNPs per sample was
21% larger in the supercentarians when
compared to average humans (p < 2.2e-
16*)
BETTER FLIES vs CONTROL Methuselah Flies vs Control: mean SNPs per
sample 23% larger(p < 2.2e-16*)
*Welch Two Sample t-test
Analysis of Supercentenarian Data
using MOSES AI toolMOSES – probabilistic evolutionary program learning algorithm from OpenCog AGI framework -- used to evolve boolean classification models
Feature set: 38k subset of SNPs in exons of genes described in: Longevity‐related molecular pathways are subject to midlife “switch” in humans.*
800 transcripts from muscle & brain tissue,
including IGF‐1/PI3K/mTOR pathway, are
coherently co-expressed only in cohort
aged 30 to 60 years
Linear model relating expression levels to
age fails in cohort aged 60 to 90.
Drugs extending lifespan in model
organisms regulate IGF‐1/PI3K/mTOR
pathway similarly to model.
MOSES results
10 cross validated runs yielded 70 models
balanced accuracy 0.73 – 1.0, F1 > 0.98
positive predictive value 1.0
Negative predictive value > 0.9986
57 SNPs in 68 HGNC genesAND
OR
OR
NOT 3:10035899
A/C
19:8547635
T/G het
17:4856312
T/C het
2061147728
T/C het
8:8099555
G/T
OpenCog-based
Gene Annotation
Service developed
by:
SingularityNET/Mozi AI Service for
Annotating/Interpreting Analysis Results
Rejuve’s mission is to accelerate longevity
therapeutics by connecting individuals, medical
clinics, biology labs and AIs.
Rejuve has been incubated since early 2018
within the SingularityNET Foundation, creators of
the leading decentralized blockchain-based AI
platform.
rejuve.io
REJUVE
REGENERATIVE
MEDICINEARTIFICIAL INTELLIGENCE
BLOCKCHAIN NETWORK
RJV tokens flow through the
network, compensating
individuals and clinics for their
data, clinics for their
therapeutic services, AIs for
their analytic services, and
researchers for their
discoveries
Biomedical researchers
push toward new
longevity therapy
discoveries via applying
SingularityNET AI to
individual members’
biomedical data
RJV: payment
for data access
RJV: payment
for data access
Data from lab experiments
AI services for analysis and
hypothesis generation
Personalized
analytics
Personal
biomedical data
RJV: payment
for data
analyticsPersonal
biomedical data
SINGULARITYNET AI AGENTSBIOMEDICAL RESEARCHERS
INDIVIDUAL MEMBERS
RJV/AGI: payment for data analytics
Biomedical researchers
push toward new longevity
therapy discoveries via
putting together data from
individual members and
member clinics
RJV: payment
for data access
New therapies
RJV/AGI: payment for AI
analytics
Data gathered during clinical
procedures
Therapies
delivered in
clinic
Data: general personal
biometric data, and data
gathered during clinical
procedures
RJV: payment
for clinical
therapeutic
procedures
Personal
biomedical data
LONGEVITY CLINICSBIOMEDICAL RESEARCHERS
INDIVIDUAL MEMBERS
SMART FASTING
APP
● Provides guidance to people wanting to
fast for health and longevity
● Gathers personal data and shapes
advice on this basis
● Data gathered goes into the Rejuve data
cloud, strongly encrypted
● Results from applying SingularityNET AI
to the data gathered are used to improve
the app’s recommendations and to fuel
longevity research
● Results and proceeds of research fueled
by user data, are shared with users