You are on page 1of 4

Brain-inspired Method for Hyper-connected and

Distributed Intelligence
Sang Ki Kim Sunme Kim
Network Research Division Network Research Division
Electronics and Telecommunication Research Institute Electronics and Telecommunication Research Institute
Daejeon, Korea Daejeon, Korea
kimsang@etri.re.kr kimsunme@etri.re.kr

Abstract — In a large-scale IoT environment such as a future created on the basis of mathematical models of the brain, and
smart city, various sensors and autonomous cars/drones/robots have recently evolved to deep learning. However, most of
collect and process a wide variety of data. However existing these machine learning methods have centralized model that
monolithic or silo style operation of smart city makes hard to concentrates data to train the model in places such as large
solve the problem when an unseen complex event occurs. In this cloud centers, and only makes an inference using derived
paper, we suggest a structure and method of distributed models at device.
intelligence processing which mimics the function and network
structure of the brain in order to efficiently control the complex This includes many problems such as network latency,
situations. Our approach utilizes local and rapid processing in bandwidth consumption, data privacy, and restrictions in
parallel as well as global and exploratory processing through information sharing between domains due to silo structure. As
cooperation and competition between local intelligence. It will a result, there is a tendency to move intelligence processing in
give us an opportunity to realize the autonomous problem the cloud into fog computing or edge computing environment.
solving on a large scale hyper-connected network However, when applying the existing centralized intelligence
infrastructure, by finding an emergence structure of distributed processing methods to these distributed computing
intelligence that mimics human consciousness process. environments directly without change, it is impossible to solve
the existing problems or provide the benefits of the distributed
Keywords— Brain Network, Automatic Thinking, Conscious
or decentralized architecture.
Thinking, Global Workspace Theory, Artificial Consciousness
To overcome those problems, we proposes a hyper-
I. INTRODUCTION connected and distributed intelligent service method which
In a future hyper-connected intelligent society where mimics the information processing structure of the brain
everyday gadgets are connected to the Internet, the composed of numerous neurons and synapses. Our approach
combination of near-infinite computing resources with a huge provides local automatic cognitive method for learning and
amount of data collected from IoT sensors will make a way to problem solving with local data and experience knowledge in
superhuman intelligence and autonomous services in all areas a distributed computing environment, as well as a global
of industry[1]. For example, numerous intelligent edge problem-solving structure and methodology with the help of
devices will interact via hyper-connected networks, and specialist agents in other domains with different knowledge
energy-constrained devices will offload complex and and cognitive functions, when encountering new problems
computationally-intensive tasks to networks with fog that cannot be solved with existing knowledge.
computing servers, infrastructure edge cloud, or in-network
computation nodes. Many distributed or decentralized II. RELATED WORK
intelligent processing nodes can create new converged The cognitive mechanism of the human brain has been the
intelligence through cooperation and competition in the form main inspiration for artificial intelligence for decades.
of a multi-agent system or swarm intelligence. Figure 1 shows Recently, as research on the structure and function of the brain
a hyper-connected and distributed intelligent service has progressed, artificial intelligence research that mimics the
environment in the future. brain brings better results. As the most of the current work in
the neural network of AI field has been constrained to emulate
the synaptic connections between neurons in the cortex of
brain, other human cognitive functions are becoming
increasingly relevant[2]. Our work is one of digital
reproduction for these human cognitive capabilities.
The architecture of human brain network also has attracted
significant attention in the research area of next generation
Internet. For example, inter-modular assortment in the
cerebral cortex of human brain has made an inspiration to the
construction model of virtual wireless sensor
network(VWSN) topology[3].

Fig. 1. Future intelligent service environment

In order to realize such a superhuman intelligence


performance, AI researchers try to utilize the state-of-the-art
human brain research. Artificial neural networks have been

978-1-7281-4985-1/20/$31.00 ©2020 IEEE 657 ICAIIC 2020


According to the recent neuroscience understanding, the elements emulating the cerebral cortex, amygdala and
topological features of human brain network support the hippocampus of the limbic system, and basal ganglia, which
segregation and integration of incoming stimuli to facilitate are the main components of the human brain.
flexible cognition and behavior[4]. These characteristics come
from the complex brain network features such as modularity,
a small-world network, and a rich-club organization[5]. In
other words, the human brain has local segregation based on
modularity as well as global integration based on a rich-club
organization as shown in figure 2.
• The local segregation model may provide a fast,
specialized, automatic, unconscious way of low-cost
thinking to a limited set of tasks on the basis of existing
learning experiences.
• The global integration model may provide a slow but
flexible and deliberate way of high-cost thinking for
forecasting, planning, and solving problems that have
not been experienced before.
Automatic thinking is processed in parallel in the
functional areas of the brain, whereas conscious thinking
involves a series of thoughts through the focus of attention. Fig. 3. Structure of distributed intelligence node

The cerebral cortex contains sensory cortex (SC) that


receives input from sensory organs, the motor cortex (MC)
that issues commands to motor organs, and association cortex
(AC) that combines inputs from different sensory organs and
controls prediction, inference, and problem-solving. The main
components for distributed intelligence are as follows;
• The sensory cortex (SC) corresponds to a functional
element that primarily processes sensing data input
from sensor devices existing in an actual external
environment.
• The motor cortex (MC) corresponds to a kind of
problem-solving engine that derives the necessary
actions with the sensing data through pattern matching.
• The basal ganglia (BG) set the priorities by analyzing
the action lists presented by the motor cortex. Herein,
Fig. 2. Cognitive functional model of the human brain the basal ganglia select an optimal alternative to return
back to the motor cortex which sends the selected
As a representative model of the global cognitive model, action as a command to an execution device in the
the global workspace theory(GWT) suggests that the outer world.
emergence of consciousness occurs through the global work
area of the cerebral association cortex. In this theory multiple, • The amygdala (AD) performs positive/negative
parallel, and specialist processes compete and co-operate for evaluations for each action from the basal ganglia by
access to a global workspace. Gaining access to the global using pre-stored information with the help of
workspace allows a winning coalition of processes to hippocampus. It may act as a kind of reward system.
broadcast information back out to the entire set of In addition, the amygdala receives reports for results
specialists[6]. We propose a distributed intelligent processing of the action executed in the real world from the
model on the basis of this theory. sensory cortex, performs the evaluation thereon, and
reflects it to the affective memory. A kind of
Further global workspace model can be applied to the reinforcement learning is done through those
study of dynamic human consciousness in order to understand procedures.
the differences between conscious and unconscious brain
events[7,8]. We may extend emergence structure of artificial • The hippocampus (HC) stores the episodic information
consciousness based on the dynamic global workspace and meta-information related to reward/punishment as
researches in the future. the evaluation results by the amygdala.

III. BRAIN-INSPIRED APPROACH TO DISTRIBUTED When there are no alternatives to an action beyond the
INTELLIGENCE
threshold in the basal ganglia, that is regarded as confronting
a problem that cannot be solved by existing problem-solving
A. Brain-inspired structure of distributed intelligence methods or new complex situations that have not been
The functional elements of the distributed intelligence experienced in the past occur. Then it is possible to try
mimicking the brain and their interactions are shown in figure problem-solving using a “conscious thinking” method based
3. The distributed intelligent node contains functional on global workspace theory through the association cortex.

658
The functional elements presented in figure 3 are The global workspace may be organized in a hierarchical
virtualization applications of a hyper-connected network structure. It may be created in each local region, and a more
infrastructure, and may be implemented, for example, in the global scale workspace is created in an upper layer thereof.
form of a virtual machine (VM), a Linux container, or a
unikernel. IV. INTEGRATED THINKING ON DISTRIBUTED
INTELLIGENCE
In addition, functional elements that are virtualization
applications may be mapped to physical devices in various A. Local automatic thinking
forms. For example, each functional element may be Figure 5 shows the processing and interaction in each
implemented as an independent physical device, as well as functional element for “local automatic thinking” at the
multiple sensory cortices and motor cortices may exist in one distributed intelligence node. The motor cortex, a problem
device as a cluster formation, and the remaining elements may solver, generates solution alternatives for a sensed situation
exist one by one. Also, it is possible to create and destroy from sensory cortex. The implementation of motor cortex is
functional elements on demand. possible by several different techniques. For example, AI
B. Global workspace model for flexible intelligence techniques such as complex event processing (CEP), artificial
neural network (ANN), and case-based reasoning (CBR) may
Global workspace is a virtual space composed of create different solutions from each other.
association cortices in multiple distributed intelligence nodes.
In order to solve new complex problems that have not been When the action alternatives are transferred to basal
experienced before, the association cortices in the global ganglia to select optimal one, the basal ganglia requests
workspace cooperate and compete. So each association cortex amygdala evaluation of alternatives in order to set the
is an intelligent agent in multi-agent system(MAS) priorities. The amygdala calculates reward point for each
environment. Figure 4 shows global workspace model which alternative with the help of hippocampus, and returns the
consists of various association cortices. results backs to basal ganglia.
Then basal ganglia sets the saliency for each alternative
and determines the priority by combining reward evaluation
of amygdala and its own policy. The action with highest
priority is delivered to motor cortex as the optimal solution for
current situation, and transformed into the actuating command
to actuator in the end.
When there are no alternatives generated by motor cortex,
in which saliency among the actions is beyond the threshold,
the problem-solving mode is changed to a “global conscious
thinking”.
The functional elements and intelligence processing
methods used for local automatic thinking act as specialists in
a specific domain. The flows may be shared because global
Fig. 4. Global workspace model with distributed intelligence nodes
conscious thinking must collectively use the problem-solving
ability of these specialists. Sub procedures I and II in the
In order to realize global workspace model, following
diagram are flows commonly used in global conscious
implementation options may be considered;
thinking.
• The mutual cooperation and competition in the global
workspace may be implemented on an existing
standard for agent communication language (ACL),
for example, foundation for intelligent physical agents
(FIPA)-ACL, or knowledge query and manipulation
language (KQML).
• The global workspace (GW) may be realized in a
pub/sub mechanism that has a relay broker or connects
all association cortices (AC) in a P2P way.
• The association cortex_sensory (ACs) and association
cortex_motor (ACm) that constructs the global
workspace are functional elements of distributed
intelligence nodes, but may also be implemented by
separating these functions from the nodes in a specific
region and gathering the same in one physical node. In
this case, the global workspace becomes a single
physical node, and the association cortex_sensory and
association cortex_motor in the node act as agents
representing each distributed intelligence node. In this
situation, it is possible to use a blackboard technique,
which is a method for distributed AI. Fig. 5. Local automatic thinking

659
B. Global conscious thinking Any distributed intelligence node can propose a problem-
Figure 6 shows the procedure when a distributed solving action for each state transition of the global workspace
intelligence node acts as a requester for global conscious using its local automatic thinking. A series of request and
thinking, while figure 7 illustrates it as a proposer of a solution. proposal on the global workspace solves the problem by
forward reasoning.
An upper flow on the figure 6 shows a step that requests
problem solving, and a lower diagram shows a flow that When the competition situation caused by multiple
determines the end of the problem solving. When the basal proposals occurs, the global workspace applies conflict
ganglia are confronted with no action alternatives beyond the resolution policy to select optimal one. The conflict resolution
threshold, it request association cortex_motor (ACm) global policy is adapted to association cortex_sensory in the form of
conscious thinking. Then ACm makes a problem specification a plug-in, thereby providing flexibility.
and distributes it to global workspace. V. CONCLUSIONS
If there is any proposal for current problem specification We propose a distributed intelligence model mimicking
on global workspace, the state of its problem solving is cognitive functions of a human brain, which integrated the
changed. When association cortex_sensory (ACs) perceives local automatic thinking and the global conscious thinking for
this change, it request motor cortex local automatic thinking a large-scale and hyper-connected IoT environment. In those
to evaluate action proposal. This triggers sub procedure 1 of situations the existing monolithic and silo model is hard to
automatic thinking in Figure 5. solve the problem when an unseen complex situation occurs.
If the result of sub procedure 1 gets the action alternative Our approach can support dynamic decision-making by
beyond threshold, global conscious thinking is completed and collective intelligence and thus provide opportunities for new
local automatic thinking continues the rest of procedure. So knowledge sharing and crowdsourcing, as if human experts
the final decision on whether or not to complete the global construct a task force to solve the hard problems.
problem solving is made by the requestor for problem solving.
Brain-inspired distributed intelligence may be integrated
into the data-centric network architecture as upper layer
functional components in the future. Advanced smart city and
connected vehicle are potential application domains to
validate our model.

ACKNOWLEDGMENT
This work was supported by the ICT R&D program of
MSICT/IITP. [2017-0-00045, Hyper-connected Intelligent
Infrastructure Technology Development].
REFERENCES

[1] Sunme Kim, Sangki Kim, Sae Hyong Park and Sang-Ha Kim, “A Data-
centric Network Architecture for Service Optimization and Data
Fig. 6. Global conscious thinking at problem solving requestor Sharing,” IEEE ICTC 2018, pp. 505–508, Oct 2018.
[2] Jesus Rodriguez, “Beyond Neurons: Five Cognitive Functions of the
Figure 7 illustrates to procedure diagrams in which a Human Brain that we are Trying to Recreate with Artificial
distributed intelligence node acts as a solution proposer of Intelligence,” https://towardsdatascience.com/ , Aug 2019.
global conscious thinking in response to the problem-solving [3] Masaya Murakami, Daichi Kominami, Kenji Leibnitz and Masayuki
requestor of figure 6. When association cortex_sensory (ACs) Murata, “Drawing Inspiration from Human Brain Networks:
perceives request of proposal on global workspace, it request Construction of Interconnected Virtual Networks,” Sensors
18(4):1133, April 2018.
motor cortex local automatic thinking to generate action
[4] Gustavo Deco, Giulio Tononi, Melanie Boly and Morten L.
proposal. Then it may make a proposal in which action Kringelbach, “Rethinking segregation and integration: contributions of
solution beyond the threshold value. whole-brain modelling,” Nature Reviews Neuroscience, Vol.16, July
2015.
[5] Alex Fornito, Andrew Zalesky and Edward T. Bullmore, Eds.
Fundamentals of Brain Network Analysis, Academic Press, 2016,
pp.1–35.
[6] Murray Shanahan, “A cognitive architecture that combines internal
simulation with a global workspace,” Consciousness and Cognition 15,
Elsvier, 2006, pp. 433–449
[7] Bernard J. Baars, Stan Franklin and Thomas Zoega Ramsoy, “Global
workspace dynamics: cortical “binding and propagation” enables
conscious contents,” Frontiers in Psychology, Vol.4, May 2013, pp. 1–
22
[8] Frank van der Velde, “In Situ representations and Access
consciousness in Neural Blackboard or Workspace Architectures,”
Frontiers in Robotics and AI, Vol.5, April 2018, pp. 1-6
Fig. 7. Global conscious thinking at solution proposer

660

You might also like