1. The Whole Brain Architecture Initiative, a specified non-profit organization
Hiroshi Yamakawa* Naoya Arakawa* Koichi Takahashi*
Reinterpreting
The Cortical Circuit
August, 19, 2017ARCHITECTURES FOR GENERALITY & AUTONOMY
http://wba-initiative.org/
npo
2. Uniformity of the neocortex is the basis for human versatility
ARCHITECTURES FOR GENERALITY & AUTONOMY
Various functions are realized on an
anatomically uniform structure
ー> general intelligence
• Bodily-kinesthetic
• Verbal-linguistic
• Logical-mathematical
• Musical-rhythmic and
harmonic
• Interpersonal
• Intrapersonal
• Visual-spatial
c.f. :
http://bio1152.nicerweb.com/Locked/media/ch48/48_
27HumanCerebralCortex.jpg
Canonical
cortical circuit
1
3. Reinterpreting the neocortex: why now?
• Importance:
The versatility of human intelligence comes from the
uniform local circuit of the neocortex.
• Feasibility:
Recent progress in neuroscience and machine
learning is making this reinterpretation more likely
• Aim:
To suggest a domain model that machine learning
experts can understand and invite them to
contribute to AGI research.
ARCHITECTURES FOR GENERALITY & AUTONOMY2
4. Can brain functions be implemented on ANNs?
① Canonical cortical circuits are gradually being unveiled
② Some cognitive functions are being implemented on
artificial neural networks (ANNs)
③ ① & ② can be combined
Yamins and DiCarlo, Nat Neurosci.19:356-65, 2016
③ Homology of CNN and neocortex for object recognition function
①
Kenneth D. Harris & Thomas D. Mrsic-Flogel, Cortical
connectivity and sensory coding, Nature 503, 51–58, 2013
ARCHITECTURES FOR GENERALITY & AUTONOMY3
5. A rough sketch of the entire brain architecture
ARCHITECTURES FOR GENERALITY & AUTONOMY
M2
M1 V1S1
EC
tactile vision audio
Hippo
campus
muscle
Environment
Basal
ganglia top-down
bottom-up
canonical cortical
circuit model
control
Reinforcement
learning
Integrating
sequential
cognition
強化学習
Lower
Higher
4
6. Reinterpretation of the cortical circuits
L1
L2/3
L4
L6
L5
State
(bottom up)
Input Output
State
(top down)
Control
(Attention,
context, etc.)
Hidden state
output
control
State
Hidden
state [Action
output]
Control
(Attention,
context, etc.)
Neocortex
Higher L2/3,
Hippocampus,
Cerebellum
Lower L2/3
(Basal ganglia-
controlled)
Thalamus
Higher L6
Lower L2/3
Basal ganglia,
[Pyramidal tracts]
Lower L6,
Lower L1
Signal Semantics Signal Semantics
Information selection
(attention)
Activitycontrol
Hidden sate
(bottom up)
(Lower L5 origin)
Thalamus
Thalamus
Higher
L2/3 or L4,
Hippocampus
Output 1
Output 2
Output 3
Input 1
Input 2
Input 3
Input 4
Input 5
ARCHITECTURES FOR GENERALITY & AUTONOMY
Signals related to the outside world: Represent beliefs and
predictions.(State = latest state, Hidden state = include history)
Other control
related signals
5
7. Bottom-up direct paths used by CNN
L1
L2/3
L4
L6
L5
State
(bottom up)
Input Output
State
(top down)
Control
(Attention,
context, etc.)
Hidden state
output
control
State
Hidden
state [Action
output]
Control
(Attention,
context, etc.)
Neocortex
Higher L2/3,
Hippocampus,
Cerebellum
Lower L2/3
(Basal ganglia-
controlled)
Thalamus
Higher L6
Lower L2/3
Basal ganglia,
[Pyramidal tracts]
Lower L6,
Lower L1
Signal Semantics Signal Semantics
Information selection
(attention)
Activitycontrol
Hidden sate
(bottom up)
(Lower L5 origin)
Thalamus
Thalamus
Higher
L2/3 or L4,
Hippocampus
Output 1
Output 2
Output 3
Input 1
Input 2
Input 3
Input 4
Input 5
ARCHITECTURES FOR GENERALITY & AUTONOMY6
8. Major pathways
• bottom-up direct path ← already described
• bottom-up indirect path
• top-down direct path #1
• top-down direct path #2
• top-down direct path #3
ARCHITECTURES FOR GENERALITY & AUTONOMY7
9. Defining the canonical cortical circuit framework
• Standard I / O semantics of neocortical local circuits
– Integrate neuroscientific findings
• Output is organized by layer
• Input is organized from the nature of source
– Semantics understandable by machine
learning specialists
• Function of neocortical local circuits
– Functions of intelligent agents to be carried out by
neocortex
• Functions suffered by damage of neocortex
• Functions that cannot bear by other brain organs
– To be consistent with existing neocortical computational
models
• E.g. visual object recognition
ARCHITECTURES FOR GENERALITY & AUTONOMY8
12. Referring various model
• Project AGI [Kowadlo 15, Kowadlo 16] was a
relatively comprehensive description
– It was consistent with many other models
– We had to adjust the input from the thalamus slightly
• The Bayesian filter has low coverage but
conservative
• In Kawaguchi hypothesis, the description of L5
was detailed and consistent with others
– There exist feed-back loops between L2 / 3 and L5
• Most of these were consistent with the
neuroscience findings of Cognitive consilience
ARCHITECTURES FOR GENERALITY & AUTONOMY11
14. Whole Brain Architecture approach @ WBAI
ARCHITECTURES FOR GENERALITY & AUTONOMY
‘to create a human-like AGI by learning from
the architecture of the entire brain’
To reach AGI, we start from mimicking high-
level architectures of the brain, and gradually
introduce details as needed
Re-
cognition
Executive
Reward
Utility
Evaluation
State
Transition
Perception Action
Intelligent Agent
Intention
Goal
State
Utility
State
State
Reward
Generation
Environment
Hippo-
campus BG
Neocortex
Amygdala
AI
(1) Develop machine
learning modules for parts
of the brain
(2) Integrate the modules to
create cogni ve
architecture
Basal Ganglia
Neocortex
Amygdala
Hippocampus
Brain
npo
13
15. Open platform strategy @ WBAI
ARCHITECTURES FOR GENERALITY & AUTONOMY
NPO WBAI is making a place for collaborative developments
Catch-up:
An open technician community that
can implement the cutting-edge
machine leanings in a short term
Integration:
A brain-inspired architecture
platform for integrating
machine learning modules
Cognitive architecture
Integrated Executive platform
Middleware for managing computational resources for executing and
learning multiple modules in parallel
ex.)BriCAcore , ROS, Flint, Nengo,
導入
ML
modules
Virtual
experiments
Sensors
Int
ro.
BriCA Language
An architecture description language that describes
connections of machine learning modules
independent of computer environment
Actuators
Whole-brain
Connectomic
Architecture
Machine
learning
repository
Github
arXiv
Catch-
up
Neuroscientific
Data
Integration
Validation
Reimplementat
ion
Improvement
Neural
Activity
Behavioral
Data
Connectome Artificial neural
network, graphical
model
npo
14
16. Open Platform Strategy @WBAI
To “harmonize AGI with human beings” by
democratizing AI technologies
( c.f. “Common goods”: Asiloma AI Principle 23 )
Let’s build a brain, together
ARCHITECTURES FOR GENERALITY & AUTONOMY
npo
15
17. Standardization trend in progress
ARCHITECTURES FOR GENERALITY & AUTONOMY
TASKS:
(1) WSJ speech corpus
(2) ImageNet dataset
(3) COCO image captioning dataset
(4) WSJ parsing dataset
(5) WMT English-German translation corpus
(6) The reverse of the above: German-English translation.
(7) WMT English-French translation corpus
(8) The reverse of the above: German-French translation.
KEY: (1) Convolutions. (2) Attention
(3) Sparsely-gated mixture-of-experts
One Model To Learn Them All
(Google, 16 Jun 2017)
16
18. Our message
Definition and implementation of the canonical neocortical
circuit is an important landmark for brain-inspired AGI.
We proposed a framework that can be understandable by
ML experts while maintaining neurobiological accuracy.
We believe that the brain-inspired approach will harmonize
AGI with mankind, and established a non-profit organization
to develop an open-platform for co-creation of brain-inspired
AGI.
If you are interested in our research and activities, please
contact us.
ARCHITECTURES FOR GENERALITY & AUTONOMY
http://wba-initiative.org/
Positions available: Posdocs/Students
17
19. We promote open co-creation of AGI on Brain
ARCHITECTURES FOR GENERALITY & AUTONOMY18
Construction of the
development environment
• World simulator
• Evaluation of AGI
• Platform for integration (BriCA)
Poster Takahashi
• Grounding to brain
Human resource
development
• Hackathon
• Seminar
• School
Research organizations
Keio University Tamagawa University
AIST AIRC
The University of
Electoro-Communications
Toshio Okubo
Kentro Goto
Ren Sakaki
Supports
Promote R&D