You are on page 1of 17

A Deep Learning Cognitive Architecture:

Towards a Unified Theory of Cognition

Isabella Panella(B) , Luca Zanotti Fragonara, and Antonios Tsourdos

Cranfield University, Cranfield MK43 0AL, UK


i.p.panella@cranfield.co.uk

Abstract. This work suggests a novel approach to autonomous systems devel-


opment linking autonomous technology to an integrated cognitive architecture
with the aim of supporting a common artificial general intelligence (AGI) devel-
opment. The paper provides a summary of strengths and weaknesses of some of
the most known cognitive architecture and highlights how to support a generic
artificial intelligent approach rather than ad hoc solutions. It also proposes objec-
tive evaluation criteria to assess a cognitive architecture. Finally, the proposed
cognitive architecture is introduced: a Deep-Learning Artificial Neural Cognitive
Architecture (D-LANCA), which aims to overcome current limits of cognitive
frameworks for autonomous systems with the view to create a common artificial
general intelligent (AGI) cognitive approach across industries.

Keywords: Cognitive architecture · Deep neural networks · Deep learning

1 Introduction
1.1 A Subsection Sample
The success of autonomous software platforms to be a ubiquitous enabling capability
delivering significant value across a range of industries and the driving requirement to
implement is dependent upon the abilities for autonomous technologies to reach a high
level of cognition in mimicking the human mind. This implies the design and imple-
mentation of computational routines able to combine cognitive abilities in an integrated
manner, what is referred to the creation of general intelligent systems [1, 2]. The first
challenge it is presented by the definition of “intelligence” within an autonomous sys-
tem. In order to design a system, the definition of what the system is needs to be clear
and associated to mathematical and physical representations. The need for the system to
be adaptable to its environment. Therefore, the author suggests the following design def-
inition for an autonomous system. Intelligence in autonomous platforms can be defined
as the ability for a system to adapt to its environment and survive. Artificial Intelligence
(AI) can be considered as a technology field that aims to embed self-modulated responses
into machines, which will enable them to be self-driven towards a primary goal: the sys-
tem survival in itself. AI can be thought as the combination of various technologies that
will provide machines with the sense of survival, hence what we described as adaptation
capabilities. Adaptation requires the machines to:

© Springer Nature Switzerland AG 2021


K. Arai et al. (Eds.): IntelliSys 2020, AISC 1250, pp. 566–582, 2021.
https://doi.org/10.1007/978-3-030-55180-3_42
A Deep Learning Cognitive Architecture 567

• Sense the environment and themselves;


• Create an inner representation of what they sense;
• Be able to reason and make inferences about the environment;
• React to the environment;
• Learn and update knowledge;
• Re-plan their course of action;
• Actuate the new plan.

For each of the points above, it is possible to associate a mathematical representa-


tion to support the implementation of each functional requirement in the design of an
intelligent system.
It is possible to distinguish two kind of “adaptability” within an autonomous system:
operational and behavioral.
Operational adaptability enables platforms to perform procedural tasks
autonomously. It refers to the “what” to perform, ranging from the sensor payload
selected for a given task to the dynamic stability of the system.
Behavioral adaptability refers to the functionalities embedded within a system that
will enable it to decide how to perform it and to adapt its behavior and reaction to the
environment it operates in. This enables the system to perform autonomously in any
environment condition, it determines the “how” a procedural task should be carried
“how” for the given situation picture.
Current approaches to autonomous systems have seen the implementation of ad
hoc and often divided approaches to this problem, with the implementation of specific
solutions to the challenge of intelligent behavior, such as planning, computer vision, lin-
guistic, problem solving, etc. This fragmented approach has led to significant advances in
computer science but has diverted the attention from the so called Unified Cognition The-
ory envisaged by Newell [1–6, 10–15], the creation of an artificial general intelligence
(AGI).
The paper is organized as following. In Sect. 2, an analysis of known cognitive archi-
tectures with their strengths and weaknesses is provided, with a summary of the gap anal-
ysis in current provided solutions. In Sect. 3, we present a suggestion on how to evaluate
objectively cognitive architectures. Section 4 presents the cognitive architecture, a Deep-
Learning Artificial Neural Cognitive Architecture (D-LANCA) that could address the
challenges identified in Sect. 2. Finally, we present conclusions and recommendations.

2 Cognitive Architecture: Strengths and Weaknesses

Cognition implies the ability of interpreting the reality and understand how things could
be in a future state to support the decision-making process. It also highlights the need
to “remember” past events to support forecasting future events outcomes. Therefore, as
reported by Vernon, D. et al. [3], cognition “allows the system to act effectively, to adapt,
and to improve”.
568 I. Panella et al.

Cognition can be considered a process through which “the system achieves


robust adaptive, anticipatory, autonomous behavior, entailing embodied perception and
action.” [3].
Cognitive architectures were proposed by Allen Newell, one of the founders of
Artificial Intelligence (AI), in 1980 in an effort to avoid “theoretical myopia” in his
search for a Unified Theory of Mind [4–6].
“Cognitive Architectures” (CA) in machines applications refer to abstract model of
cognition in artificial agents with the associated software functionalities linked to their
implementation in autonomous systems through Artificial Intelligence (AI) methodolo-
gies. As stated by Langley et al. [6], they can be considered “the underlying infrastructure
for an intelligent system”. Profanter, S. [7], defines a cognitive architecture as “a blueprint
for an intelligent system”, whereby “blueprint” refers to the design characteristics of the
framework, such as the assumptions made to support representation, its memory char-
acteristics, and the process used to operate those memories. Their main role is to realize
artificial systems that exhibit intelligent behaviors in general setting mimicking human
behavior [7]. The objective is to model complex and higher-level functionalities, such
as language, reasoning, problem solving, decision making, or planning, which involve
complex knowledge structures and complex information extraction methods [8].
Extensive surveys of cognitive architectures have been carried out by [6, 8, 13,
17], and [13] among others. A summary of the most adopted cognitive architectures’
implementation has been presented by the authors in the table below (Table 1). Herein,
the architectural approach has been highlighted as well as the learning algorithms. The
rational for highlighting the learning algorithms is that they represent the ability of
the system to behaviorally adapt to the environment. In the table, we also highlight
the architecture characteristics and we then assess the weakness that each architecture
presents with respect to its application to autonomous platforms.

2.1 Gap Analysis of Current Cognitive Architectures

It is possible to observe from the analysis in Table 1 that whilst significant achievements
have been accomplished over time by the various cognitive architectures’ developments,
there are still open research questions and challenges related to:

• Integration of distinct cognitive functions;


• Computational efficiency of the platform;
• Parallel processing of data - even though current architectures have some commitment
to parallelism, especially in memory retrieval, they tend to rely on one or few decision
modules [6];
• Integration of heterogeneous knowledge representation;
• Integration of heterogenous reasoning and cognitive functions;
• Integration of planning, acting, monitoring and goal reasoning;
• Knowledge problem - knowledge acquisition, knowledge size, and homogeneity and
homogeneous typology of encoded knowledge used by the agents in reasoning, deci-
sion making, learning, planning, etc. [20, 22]. Knowledge acquisition is the ability
A Deep Learning Cognitive Architecture 569

Table 1. Cognitive architectures characteristics and weaknesses [8–23]

Approach Learning algorithms Characteristics Weaknesses


CLARION Artificial General Reinforcement The working memory is a No general and
– Connectionist Intelligence (AGI) Learning separate structure cross-domain
Learning with – software agent-based Hebbian Learning Semantic, episodic, and knowledge
Adaptive Rule architecture Bayesian update procedural memory are – difficulties in
Induction On-line Knowledge and Gradient descent both implicit and explicit dealing with variety of
experiences are methods (chunks/rules, NN) data coming from
represented in CLARION Learning of new It has a reward system in different sources
within an implicit-explicit representations the form of a motivational The dual-layered
dichotomy: using chunks through new chunks, subsystem and a conceptual
and rules (for explicit new rules, and new meta-cognitive subsystem information does not
knowledge), and neural NN representations (MCS determines rewards provide the possibility
networks (for implicit Hypothesis testing based on the MS) of encoding (manually
knowledge) rule learning for Support autonomous or automatically via
It is based on a four-way explicit learning learning from new learning cycles) the
division of implicit versus Bottom-up rule representation information in terms
explicit knowledge and learning – implicit of the heterogeneous
procedural versus to explicit learning classes of
declarative knowledge representations
It addresses both top-down Possible co-existence
learning (from explicit to of different
implicit knowledge) and representations of
bottom-up learning (from prototypes, exemplars
implicit to explicit and theories (and the
knowledge) interaction among
It is an hybrid paradigm them) is not addressed
architecture
SOAR – State, Artificial General Reinforcement Problem Space Hypothesis One decision at the
Operator and Intelligence (AGI) Learning framework time – single operator
Result – software agent-based Chunking to form Working, semantic, and can be selected at
architecture new rules episodic memory each step, forcing a
Rule based architecture – production rules – Relational Graph serial bottleneck
designed to model general (procedural structure: Knowledge is Impasse – when
intelligence long-term memory) represented as rule knowledge about
Knowledge and Bayesian update organized operators operator selection is
experiences are Complex rule sets in Procedural memory – Rule insufficient or when
represented using rules planning, problem Iconic memory explicitly an abstract operator
(procedural knowledge), solving and natural defined cannot be
relational graph structure language Reward system – appraisal implemented
(semantic knowledge), comprehension in based as well as user Goals are
and episodes of relational real-time distributed defined internal/external hierarchically
graph structures (episodic environments reward organized
memory) Experience based Specific modalities: The design of the
It is a symbolic learning Visual input perceptual-motor
architecture Auditory Input systems within SOAR
is unrealistic,
requiring users to
define their own input
and output functions
for a given domain
Approximate
comparisons are hard
and computationally
intensive as
implemented through
graph-like
representations
(continued)
570 I. Panella et al.

Table 1. (continued)

Approach Learning algorithms Characteristics Weaknesses


SOAR Agents are not
endowed with general
knowledge and only
process ad-hoc built
(or task-specific
learned) symbolic
knowledge structures
ACT-R (Adaptive Biologically based Reinforcement Supports semantic memory It cannot learn in real
Control of cognitive architecture to learning for (encoded as chunks) and time
Thought model human Behavior productions (linear episodic memory It cannot learn from
– Rational) Knowledge and discount version) Buffers encode working arbitrary stationary
experience are represented Bayesian update for and episodic memory large and
using chunks and memory retrieval Specific modalities: non-stationary
productions Production rules Visual input databases – lack of
It is a hybrid paradigm generation for Auditory Input adaptiveness to the
architecture learning of new Basic motor functions environment
representations ACT-R allows to represent Lack of goal
the information in terms of prioritisation
prototypes and exemplars Lack of Adaptive
and allow to perform, heterogeneous fusion
selectively, either prototype Lack of object feature
or exemplar-based search in an
categorization. This means environment
that this architecture allows Does not assume a
the modeler to manually heterogeneous
specify which kind of perspective – Cannot
categorization strategy to deal with conflicting
employ according to his information
specific needs Cannot manage
different reasoning
strategies
It is not able to
autonomously decide
which reasoning
procedures to activate
Task-specific
knowledge and not
with general
cross-domain
knowledge
NARS Artificial General Unified reasoning Inference Engine Task-specific
– Non-Axiomatic Intelligence (AGI) mechanism on a Integrated Memory knowledge and not
Reasoning – software agent-based unified memory for Control Mechanism with general
System architecture learning, Knowledge representation cross-domain
Knowledge and with an experience knowledge
experience are represented grounded semantics of the Lack of knowledge
using beliefs, tasks, and language, a set of inference heterogeneity
concepts rules with non-axiomatic structures – lack of
It is a symbolic logic support adaption in interaction between
architecture case of insufficient different reasoning
knowledge and resources strategies
The whole memory is
semantic and incorporates
cognitive maps
(continued)
A Deep Learning Cognitive Architecture 571

Table 1. (continued)

Approach Learning algorithms Characteristics Weaknesses


Strong in reasoning with
insufficient knowledge and
resources
It adopts a unified
reasoning mechanism on a
unified memory
LIDA – Learning Artificial General Constraint Cognitive cycle Task-specific
Intelligent Intelligence (AGI) satisfaction (action-perception cycle) knowledge and not
Distribution – software agent-based Global Workspace acting as cognitive atom with general
Agent architecture Theory paradigms Cognitive cycles include cross-domain
Knowledge and Reinforcement sensory, perceptive, knowledge
experience are represented learning associative, workspace, Lack of knowledge
using perceptual Global broadcasting transient episodic, heterogeneity
knowledge – nodes and declarative, procedural, structures – lack of
links in a Slipnet-like net global workspace, and interaction between
with sensory data of sensory motor memory different reasoning
various types attached to Distinct modules for strategies
nodes perception, each of the
Episodic knowledge listed memories, action
– Boolean vectors (Sparse selection,
Distributed Memory It can support control
Procedural knowledge structures for software
– schemes a la Schema agents and robots
Mechanism It possesses an explicit
It is an hybrid paradigm attention mechanism [18,
architecture 19]
ART (Adaptive Biologically based Non-linear neural Supports working memory Computationally
Resonance cognitive architecture to networks with (LTM Invariance Principle intensive
Theory) model human Behavior feedback and Inhibition of Return Lack of knowledge
Tries to model human – categorize events rehearsal law), semantic heterogeneity
memory and Unsupervised memory (Limited structures – lack of
consciousness learning association between interaction between
Knowledge and – categorization chunks), episodic memory different reasoning
experience are represented Supervise learning (limited spatial and strategies
using visual 3D boundary – anomaly detection temporal representations),
and surface representation Real-time learning procedural memory
It is a hybrid paradigm – match learning and (multiple explicitly defined
architecture Vector Associative neural systems for
Map (VAM) learning, planning and
Reinforcement control of action), iconic
learning – model memory (emerges from
how amygdala and role of top-down attentive
basal ganglia interactions in laminar
interact with models of how the visual
orbitofrontal cortex cortex sees), perceptual
Bayesian effects as memory (model
emergent properties development of laminar
Hebbian and visual cortex), cognitive
non-Hebbian map (neural networks),
properties in reward system (model how
learning dynamics amygdala, hypothalamus,
Self-organizing and basal ganglia interact
maps with gradient with sensory and prefrontal
descent – learning of cortex to learn to direct
new representations attention and actions
towards valued goals), and
attention control and
consciousness
(continued)
572 I. Panella et al.

Table 1. (continued)

Approach Learning algorithms Characteristics Weaknesses


ART predicts a link
between processes of
Consciousness, Learning,
Expectation, Attention,
Resonance, and Synchrony
(CLEARS)
Learning from arbitrarily
large databases
The neural network
enables local computations
ART models can
autonomously learn from
non-stationary
environments
Memory stability
– matches bottom up and
top down representations
Heterogeneous sensor
fusion through
multi-modal feature and
hierarchical rule
combinations
Brain imaging studies
CoJACK Knowledge and Reinforcement It includes sets of beliefs Time cost associated
(Cognitive Java experiences are learning for long term memory with adding beliefs or
Agent represented using Bayesian Update Procedural memory instantiating a plan
Construction Kit) Beliefs-Desires-Intentions BDI addresses events, plans, Add noise to the
(BDI) architecture that and intentions decision-making
handles events, plans and Event and goal manager process, which can
intentions (procedural are included affect the retrieval of
memory), belief sets It can function beliefs and affect the
(declarative memory) and autonomously selection of next
activation levels It uses ACT-R declarative intention to execute
It has been used to model memory equations
the variation in human It gets input from the world
Behavior as events, which are then
It is a hybrid paradigm processed by plans
architecture
ICARUS Integrated cognitive Reinforcement It includes several Lack of concurrent
architecture for physical learning modules: processing to cope
agents with knowledge State action pairs Perceptual system with asynchronous
specified in the form of Belief desired and Planning system inputs from multiple
reactive skills, each intention agents Execution system sensors while
denoting goal with high utility Several memory systems coordinating
It is a symbolic selection criteria Concepts are matched to resources and actions
architecture Search trees percepts in a bottom-up across different
way and goals are matched modalities
to skills in a top-down way Uncertainty is not
Conceptual memory addressed
contains knowledge about
general classes of objects
and their relationships,
while skill memory stores
knowledge about the ways
of doing things
It has a
Long-Term-Memory
(LTM) and a Short-Term
Memory (STM)
A Deep Learning Cognitive Architecture 573

to dynamically and efficiently store and retrieve knowledge based on the experiences
or events encountered by the system within the environment. Knowledge size refers
to the dimension of the knowledge base available to the agents. Knowledge typology
refers to theories on how humans organize, reason, and retrieve conceptual informa-
tion. In the past, it was believed that concepts representation in the human brain was
homogeneous and that concepts were categorized as classical, a prototype view, exem-
plar view, or theory-theoretical view. However, it is now believed that human may use
in different instances different representation to categorize concepts, which led to the
Heterogeneous Hypothesis about the nature of concepts. The heterogeneous hypoth-
esis claims that different type of representation may exist and perhaps co-exist within
the human brain. All such representations constitute different body of knowledge and
contain different type of information associated with the same type of entity. More-
over, the heterogeneous hypothesis claims that each body of conceptual knowledge is
distinguished by specific processes in which such representations are involved, such
as in tasks like recognition, learning, categorization, planning, etc. The heterogeneous
hypothesis, which assumes the availability of different types of knowledge encoded
in a conceptual structure, is not implemented in any CA [20].

3 Cognitive Architecture: Evaluation Criteria

An additional challenge linked to the implementation of cognitive architectures is rep-


resented by the identification of key performance criteria to support their evaluation.
Anderson and Lebeire [14] proposed to use the 12 criteria identified by Newell in 1980
to evaluate cognitive architectures on how well they do meet these functional criteria.
These criteria represent an attempt for Newell to focus the field of cognitive architec-
tures on the big picture needed to understand the human mind. They suggest calling the
evaluation of the theory by this set of criteria “The Newell Test”.
The criteria are reported in Table 2, first column. The first nine criteria reflect things
that the architecture must achieve in order to implement human intellectual capability.
The last three reflect constraints on how these functions must be achieved. In the second
column of the table, the author has associated the artificial intelligence methodologies
that could be used to assess a software agent architecture. In the third column, the author
has created metrics associated to the functional criteria to objectively evaluate cognitive
architectures criteria in a software agent platform.
The importance of considering quantitative evaluation criteria for cognitive archi-
tecture is to ensure that they can be applied across a range of engineering and industrial
applications, rather than limited to studies on human cognition, behavior, and human
performance. However, in the research the author carried out, cognitive architectures
seem to be almost ignored in engineering autonomous vehicles application design.
574 I. Panella et al.

Table 2. Newell’s functional criteria for human cognitive architectures reported in Table 1 in [14]
with associated AI technologies and Metrics for CA evaluation generated by the author.

Criteria AI technologies to realize the Metrics to objectively evaluate


functionality CA
Functional criteria
Flexible Behavior – Behave Learning Percentage in Operational
flexibly as a function of the Symbolic reasoning – It Performance Accuracy
environment enables to perform arbitrary – Robustness in Operational
task to high level of expertise changes in tasks
– cognitive plasticity Percentage of Errors in
Rule based systems completing a task as it should
Learning rate
Real-time performance Neural network – parallel Computational time to carry
– Operate in real-time processing out a task
It becomes a constraint on
learning as well as
performance
Adaptive Behavior – Exhibit Decision Making – Decision Percentage of action selection
rational, that is, effective trees and Utility function resulting in positive task
adaptive Behavior – Does the maximization completion
system yield functional Errors in action selected vs
Behavior in the real world? outcomes
Robustness in Action Selection
– Repeatability of right action
selection to complete a task as
it should – error in action
selected can be considered as a
limit that tends to zero
Vast Knowledge Base – Vast Expert systems Minimum data required to
knowledge of the Big Data technologies, perform tasks in a given
environment Data warehousing – Create environment – Assess
heterogeneous classes of data Architecture performance in
to use as ground truth for completing tasks vs. amount of
computational functions data available and stored in
databases
Percentage of Errors in
performing tasks vs richness
of database
Computational Memory
capacity required by the
system
(continued)
A Deep Learning Cognitive Architecture 575

Table 2. (continued)

Criteria AI technologies to realize the Metrics to objectively evaluate


functionality CA
Dynamic Behavior – Behave Information fusion, Elapsed Time between
robustly in the face of error, Reasoning reasoning and action selection
the unexpected, and the Decision making to support states changes in
unknown. The ability to deal Probabilistic model – Bayes response to external inputs
with a dynamic and networks Number of False positive and
unpredictable environment is False Negatives
a precondition to survival for
all organisms
Knowledge Integration Knowledge Base Systems Number of data classes in
– Integrate diverse Inference current databases – Update of
knowledge: Symbols and Utility functions knowledge databases with
abstraction Logic information heterogeneously
sourced
Computational Memory
capacity required by the
system
Natural Language Symbolic reasoning Voice commands – Ability to
Speech to text technologies control the system through
Classification and clustering natural language interaction
Consciousness – Exhibit Anomaly Detection Percentage of Forecast
self-awareness and a sense of Meta-reasoning Algorithms False positive and
self Diagnostic False Negatives – System
Prognostic health monitoring – diagnostic
Forecasting and prognostics
Behavior Robustness (%)
Behavior monitoring
– feedback on actions vs.
results – if the action produces
the wanted result/task
completion, then the behavior
is good. Measure percentage
of accomplishing the task
Percentage of reasoning time
per task completion
– Feedback on reasoning
– reasoning technologies
assessed against situation
picture
Computational Memory
capacity required by the
system
(continued)
576 I. Panella et al.

Table 2. (continued)

Criteria AI technologies to realize the Metrics to objectively evaluate


functionality CA
Learning – Learn from its Learning Computational Memory
environment Short and Long-term memory capacity required by the
logical design system
Inference engine Rate of learning in the system
Anomaly detection
Rule based systems
Functional constraints criteria
Development – Acquire Learning Computational Memory
capabilities through capacity required by the
development system
Percentage of Knowledge
database reusability – Ability
to adapt knowledge acquired
to different scenarios
Evolution – Arise through Evolutionary Algorithms Scalability – Software
evolution – Does the theory Reinforcement learning Functions self-building
relate to evolutionary and Maintainability – Software
comparative considerations? Functions self-upgrades
Brain – Be realizable within Integrated knowledge base Overall system performance
the brain: Do the components algorithms
of the theory exhaustively
map onto brain processes?

4 Deep-Learning Artificial Neural Cognitive Architecture


(D-LANCA)
The aim of this section is to address the current limits of cognitive architectures for
autonomous systems specifically the issue of parallel execution of functionalities, hetero-
geneous knowledge representation, whilst improving computational efficiency, minimiz-
ing the limitations of historic knowledge database and knowledge limit, whilst improving
real-time decision-making abilities. Specifically, the author is focusing on addressing
the heterogeneity problem, which assumes the existence of multiple representations for
a concept, and each concept represented by different kinds of categorization and rea-
soning mechanisms and processes are assumed to exist and require integration within a
common framework. The objective is to create a common artificial general intelligent
(AGI) cognitive architecture that can enable multi-functional processing and real time
learning for autonomous driving systems.
The cognitive architecture is going to be developed through a modified deep learning
neural network (DLNN) framework, which will enable the parallel processing of infor-
mation and will be developed as a multi-agent systems (MAS) framework, interfacing
with a knowledge base system represented by short and long-term memories, as well as
A Deep Learning Cognitive Architecture 577

a goal management system to support indirect communication among nodes in case of


lack of connections.
The rational to adopt a deep learning framework to implement a cognitive architecture
resides in the ability of deep learning to support computational models with multiple
process layers and learn representation of data with multiple level of abstraction [23].
Deep learning has been successfully applied in speech recognition, visual object
recognition [23–29], object detection, drug discovery, genomics, etc. It enables the dis-
covery of “intricate structure in large data sets by using the backpropagation algorithm
to indicate how a machine should change its internal parameters that are used to compute
the representation in each layer from the representation in the previous layer. Deep learn-
ing enables the computer to learn anything without human intervention and differs from
traditional AI techniques as it enables learning though a hierarchical model in which
each layer represents a level of abstraction for the problem versus the feature engineer-
ing, i.e. the manual definition of features within a data set. Also, Deep Learning works
with unlabeled data and, most importantly, enables multiple decisions simultaneously.
The issue of dimensionality and scalability, as well as the knowledge size challenge
previously described, will be addressed by automatically adding or removing nodes and
connections within the network, or the self-programming characteristics of ANN (Rizk
et al. 2019). This will enable to model the brain neuroplasticity. Plasticity can be advan-
tageous to reduce computational time by reducing the search space to support decision
making, actuation, and sensor selection, supporting efficiency in power management and
goal management. Multiple goals can be passed to the system as an additional neural
network layer. Plasticity can ease handling very large, hybrid, knowledge spaces and
selection of actions to support dynamic system adaption and reaction in as real-time
environment. The architecture has been named D-LANCA (Deep-Learning Artificial
Neural Cognitive Architecture).
The cognitive process in D-LANCA is based on Boyd’s OODA (Observe-Orient-
Decide-Act) loop [31] and [32].
The need to associate a cognitive architecture to the OODA loop is funded on the
need to associate operational and behavioral autonomous technologies within a unique
framework.
The association of sensing functional elements within the observe class, the reasoning
association within the orient class, and the decision making within the decide class of
the OODA loop enable the autonomous technologies to be mapped onto an operational
framework for the implementation of dynamic real-time autonomous platforms.
It is now important to create a link between the OODA loop and the Deep Learning
Neural Network system as D-LANCA is designed as a modified deep neural network
framework. The D-LANCA cognitive architecture is described in Fig. 1 and Fig. 2, in
which the authors highlights the required changes.
Specifically, the modifications proposed to a conventional deep learning neural
network framework are the following and captured in Fig. 3:

1. The propagation function in the neural layer ( ) is no longer represented by a
weighted sum of the input multiplied by their weights, but it becomes an opera-
tor node within the framework, whereby sensor fusion, information extraction and
optimization, as well as the vehicle motion control can be implemented.
578 I. Panella et al.

Fig. 1. D-LANCA linked to OODA loop

2. The weights modification in the network can be represented through the implemen-
tation principles of a spiking neural network (SNN), which will enable a bio-inspired
learning through the weight modification based on the temporal and performance
information provided by each modified activation function [27, 32–41]. By com-
puting the weights through the temporal contribution of each activation function we
can surpass the issues of training the SNN, as they are implemented through sums
of Dirac delta functions that do not have derivatives to support a backpropagation
algorithm implementation to test the network.

Fig. 2. D-LANCA architecture.

3. The activation function in the neural network is now considered as multi-agent soft-
ware function, whereby computations such as information extraction, learning with
for instance reinforcement learning (RL), decision making, etc. will be performed.
By doing so, the software cognitive functionalities do not require to be executed
A Deep Learning Cognitive Architecture 579

sequentially but they can be processed in parallel and the propagation function will
output a multi-dimensional array to determine which functions need to be enabled.
4. The bias in the neural network are now considered inputs from knowledge databases,
goals, human input, and world model created within the system.
5. As reported in [46], subjective estimation of the environment enables to create mean-
ing within the system. However, subjective bias cannot be constants and they need
to evolve as a function of the increased knowledge of the system. Therefore, as the
databases are updated, the meaning of the semantic knowledge of the system will
change and the bias will support the improvement in the overall inference of the
system.
6. Learning, represented by the adaption of the network to better handle the task at
hand within the given environment, is no longer solely represented by the adaption
of the weights within the system, but by the expansion and contraction of the net-
work through the inhibition of specific agents within the activation function as not
relevant for the task, for instance, or the addition of new nodes to include additional
information or inputs. This enables the implementation of plasticity as an embedded
property within the system to learn the neural network structure and enable adaptive
structures. In Russell and Norvig [30], one of the major problems in modelling a
neural network is identified to be overfitting, which occurs when there are too many
parameters in the model. Where when dealing with deep neural networks, one of the
challenges in modelling the problem is represented by the number of hidden layers
needed to support its solution and their sizes. A potential solution to plasticity is
provided by Russell and Norvig ([30], p. 748) with the introduction of the optimal
brain damage algorithm, which starts with a fully connected neural network and
then removes connections from it, and with the tiling algorithm, which, on the other
hand, starts with a single node network (perceptron) that tries to produce the correct
output for as many example training sets as possible and adds subsequent nodes to
handle the ones that could not be handled by the single perceptron.

The deep neural network is a recurrent neural network with the feedback loops
supporting the implementation of cognition cycle.

Fig. 3. D-LANCA explicit link to Deep Neural Network Architecture.


580 I. Panella et al.

5 Conclusions and Further Work


In the work herein presented, the author has provided a systemic approach to support
the design of a cognitive architecture able to support artificial general intelligence (AGI)
implementation for an autonomous system.
An objective evaluation of commonly used cognitive architectures has been presented
and metrics to evaluate their performance suggested.
The work has presented a deep neural network modified architecture as a cognitive
architecture (D-LANCA). One of the advantages of using D-LANCA is the ability to
support a robust representation of knowledge and support parallel functionalities to be
co-processed in real time, mimicking the structure of a human brain. In addition, D-
LANCA is envisaged to minimize the challenges of working in partially observable or
not observable environment to support increased trust in the machine ability to learn
and make decisions. In D-LANCA the author suggests to use the databases of historic
data on behavior, environmental conditions, performance limits as a bias input within the
system to support reasoning under uncertainty without limiting the learning and decision
making ability of the machine
Further work is currently carried out to derive the mathematical model of D-LANCA
and to implement a simulation model to test its performance against the derived eval-
uation criteria presented in Sect. 3. The author is focusing on the demonstrating the
feasibility of changing the activation function into a MAS processing unit by imple-
menting a reinforcement learning (RL) to create a trajectory planning function within
the decision stage in the cognitive cycle. The rational for adopting RL as one of the func-
tional building blocks in the activation function is that enables a fully data driven and
self-learning model that does not rely on predictions, predefined rules or prior human
knowledge [28]. This will address the challenge in AI of making good sequences of
decisions under uncertainty. It will be assumed that information is readily available for
the decision-making stage to be computed.
The applicability of this method will be demonstrated via modification of deep neural
network focusing on the Decision Making/Thinking subsystem by deriving the equations
to describe this subsystem and implementing a Matlab/Simulink modelling to support
its analysis and evaluation.

References
1. Langley, P.: Cognitive architectures and general intelligent systems. AI Mag. 27(2), 33–44
(2006)
2. Langley, P.: Information-processing psychology, artificial intelligence, and the cognitive
systems paradigm thanks to. In: AAAI (2017)
3. Vernon, D., Metta, G., Sandini, G.: A survey of artificial cognitive systems: implications for
the autonomous development of mental capabilities in computational agents. IEEE Trans.
Evol. Comput. 11(2), 151–180 (2007). https://doi.org/10.1109/TEVC.2006.890274
4. Models, C., Branch, A., Force, A., Patterson, W., Force, A.: Unified Theories of Cognition:
Newell’s Vision after 25 Years Presenters, pp. 250–251 (2012)
5. Anderson, J.R., Bothell, D., Byrne, M.D., Douglass, S., Lebiere, C., Qin, Y.: An integrated
theory of the mind. 111(4), 1036–1060 (2004). https://doi.org/10.1037/0033295x.111.4.1036
A Deep Learning Cognitive Architecture 581

6. Sun, R., Langley, P., Laird, J.E., Rogers, S.: Cognitive architectures: research issues and
challenges. Cogn. Syst. Res. 10(2), 141–160 (2009). https://doi.org/10.1016/j.cogsys.2006.
07.004
7. Profanter, S.: Cognitive architectures (2012)
8. Lieto, A., Bhatt, M., Oltramari, A., Vernon, D.: The role of cognitive architectures in general
artificial intelligence. Cogn. Syst. Res. 48, 1–3 (2018). https://doi.org/10.1016/j.cogsys.2017.
08.003
9. Duch, W., Oentaryo, R.J., Pasquier, M.: Cognitive architectures: where do we go from here?
Front. Artif. Intell. Appl. 171, 122–136 (2008)
10. Thagard, P.W.: Cognitive architectures. In: The Cambridge Handbook of Cognitive Science.
Cambridge University Press, pp. 50–70 (2012)
11. Ritter, F.E.: Two cognitive modeling frontiers. Trans. Jpn. Soc. Artif. Intell. 24, 241–249
(2009). https://doi.org/10.1527/tjsai.24.241
12. Kotseruba, I., Tsotsos, J.K.: A Review of 40 Years of Cognitive Architecture Research: Core
Cognitive Abilities and Practical Applications (2016)
13. Ye, P., Wang, T., Wang, F.Y.: A survey of cognitive architectures in the past 20 years. IEEE
Trans. Cybern. 48(12), 3280–3290 (2018). https://doi.org/10.1109/TCYB.2018.2857704
14. Anderson, J.R., Lebiere, C.: The Newell Test for a Theory of cognition
15. Samsonovich, A.: Comparative Table of Cognitive Architectures (started on October 27, 2009;
last update: June 18, 2012)
16. Samsonovich, A.V.: Comparative analysis of implemented cognitive architectures. Front.
Artif. Intell. Appl. 233, 469–479 (2011). https://doi.org/10.3233/978-1-60750-959-2-469
17. Kingdon, R.: A review of cognitive architectures. ISO Project report (2008)
18. Franklin, S., Madl, T., D’Mello, S., Snaider, J.: LIDA: a systems-level architecture for cogni-
tion, emotion, and learning. IEEE Trans. Auton. Ment. Dev. 6(1), 19–41 (2014). https://doi.
org/10.1109/TAMD.2013.2277589
19. Computing, C.: The Mind According to LIDA - A Brief account The “LIDA Model” and its
Cognitive Cycle, pp. 1–20 (2013)
20. Lieto, A., Lebiere, C., Oltramari, A.: The knowledge level in cognitive architectures: current
limitations and possible developments. Cogn. Syst. Res. 48, 39–55 (2018). https://doi.org/10.
1016/j.cogsys.2017.05.001
21. Li, D.: A tutorial survey of architectures, algorithms. APSIPA Trans. Signal Inf. Process.
3(2014), 1–29 (2014)
22. Lieto, A.: Representational limits in cognitive architectures. CEUR Workshop Proceedings,
vol. 1855, pp. 16–20 (2017)
23. Lecun, Y., Bengio, Y., Hinton, G.: Deep learning. Nature 521(7553), 436–444 (2015)
24. Liu, W., Wang, Z., Liu, X., Zeng, N., Liu, Y., Alsaadi, F.E.: Neural networks architectures
review. 1–31 (2017)
25. Liu, Y., Xiang, C.: Hybrid learning network: a novel architecture for fast learning. Procedia
Comput. Sci. 122, 622–628 (2017)
26. Luo, X., Shen, R., Hu, J., Deng, J., Hu, L., Guan, Q.: A deep convolution neural network model
for vehicle recognition and face recognition. Procedia Comput. Sci. 107(ICICT), 715–720
(2017)
27. Petersen, S.E., Sporns, O.: Brain networks and cognitive architectures. Neuron 88(1), 207–219
(2015)
28. Qi, X., Luo, Y., Wu, G., Boriboonsomsin, K., Barth, M.: Deep reinforcement learning enabled
self-learning control for energy efficient driving. Transp. Res. Part C Emerging Technol. 99,
67–81 (2019)
29. Rizk, Y., Hajj, N., Mitri, N., Awad, M.: Deep belief networks and cortical algorithms: a
comparative study for supervised classification. Appl. Comput. Inf. 15(2), 81–93 (2019)
582 I. Panella et al.

30. Russell, S.J., Norvig, P.: Artificial intelligence: a modern approach, vol. 9 (1995)
31. Behere, S., Törngren, M., Chen, D.: A reference architecture for cooperative driving. J. Syst.
Architect. 59(10), 1095–1112 (2013)
32. Brehmer, B.: The dynamic OODA loop: amalgamating Boyd’s OODA loop and the cybernetic
approach to command and control. In: 10th International Command and Control Research
and Technology Symposium The Future of C2 (2005)
33. Huyck, C.R.: A neural cognitive architecture. Cogn. Syst. Res. 59, 171–178 (2020)
34. Kim, J., Kim, H., Huh, S., Lee, J., Choi, K.: Deep neural networks with weighted spikes.
Neurocomputing 311, 373–386 (2018)
35. Sboev, A., Vlasov, D., Rybka, R., Serenko, A.: Spiking neural network reinforcement learning
method based on temporal coding and STDP. Procedia Comput. Sci. 145, 458–463 (2018)
36. Tavanaei, A., Ghodrati, M., Kheradpisheh, S.R., Masquelier, T., Maida, A.: Deep learning in
spiking neural networks. Neural Netw. 111, 47–63 (2019)
37. Wu, X., Wang, Y., Tang, H., Yan, R.: A structure-time parallel implementation of spike-based
deep learning. Neural Netw. 113, 72–78 (2019)
38. Wang, B., Chen, L.L., Zhang, Z.Y.: A novel method on the edge detection of infrared image.
Optik 180, 610–614 (2019)
39. Stief, P., Dantan, J.-Y., Etienne, A., Siadat, A.: A New Methodology to Analyze the Functional
and Physical Architecture of Existing Products for an Assembly Oriented Product Family
Identification (2018)
40. Seijen, V., Harm, M.F., Romoff, J., Laroche, R., Barnes, T., Tsang, J.: Hybrid reward archi-
tecture for reinforcement learning. In Advances in Neural Information Processing Systems
2017 (NIPS 2017), pp. 5393–5403 (2017)
41. Qi, X., Luo, Y., Wu, G., Boriboonsomsin, K., Barth, M.: Deep reinforcement learning enabled
self-learning control for energy efficient driving. Transp. Res. Part C Emerging Technol. 99,
67–81 (2019)

You might also like