COMPUTATIONAL INTUITION INTO POTENTIAL COGNITIVE ARCHITECTURES

Cognitive architectures are a part of research in general AI, which began in the 1950s with the goal of creating programs that could reason about problems across different domains, develop insights, adapt to new situations, and reflect on themselves. Similarly, the ultimate goal of research in cognitive architecture is to model the human mind, which will bring us closer to building human-level artificial intelligence. Cognitive architectures attempt to provide evidence that particular mechanisms succeed in producing intelligent behavior and thus contribute to cognitive science. Moreover, the body of work represented by the cognitive architectures, and this review, documents what methods or strategies have been tried previously, how they have been used, and what level of success has been achieved or lessons learned, all important elements that help guide future research efforts. For AI and engineering, documentation of past mechanistic work has obvious import. But this is just as important for cognitive science since most experimental work eventually attempts to connect to explanations of how observed human behavior may be generated and the body of cognitive architectures provides a very rich source of viable ideas and mechanisms.

LEADING-EDGE OF COGNITIVE ARCHITECTURE

According to Russel and Norvig artificial intelligence may be realized in four different ways: systems that think like humans, systems that think rationally, systems that act like humans, and systems that act rationally. The existing cognitive architectures have explored all four possibilities. For instance, human-like thought is pursued by the architectures stemming from cognitive modeling. In this case, the errors made by an intelligent system should match the errors typically made by people in similar situations. This is, in contrast, to rationally thinking systems which are required to produce consistent and correct conclusions for arbitrary tasks. A similar distinction is made for machines that act like humans or act rationally. Machines in either of these groups are not expected to think like humans, only their actions or behavior is taken into account.

However, with no clear definition and the general theory of cognition, each architecture was based on a different set of premises and assumptions, making comparison and evaluation difficult. Several papers were published to resolve the uncertainties, the most prominent being Sun’s desiderata for cognitive architectures and Newell’s functional criteria. Newell’s criteria include flexible behavior, real-time operation, rationality, large knowledge base, learning, development, linguistic abilities, self-awareness, and brain realization. Sun’s desiderata are broader and include ecological, cognitive, and bio-evolutionary realism, adaptation, modularity, routineness, and synergistic interaction. Besides defining these criteria and applying them to a range of cognitive architectures, Sun also pointed out the lack of clearly defined cognitive assumptions and methodological approaches, which hinder progress in studying intelligence. He also noted an uncertainty regarding essential dichotomies, modularity of cognition, and structure of memory. However, a quick look at the existing cognitive architectures reveals persisting disagreements in terms of their research goals, structure, operation, and application. Instead of looking for a particular definition of intelligence, it may be more practical to define it as a set of competencies and behaviors demonstrated by the system. While no comprehensive list of capabilities required for intelligence exists, several broad areas have been identified that may serve as guidance for ongoing work in the cognitive architecture domain. For example, Adams et al. suggest areas such as perception, memory, attention, actuation, social interaction, planning, motivation, emotion, etc. These are further split into subareas. Arguably, some of these categories may seem more important than others and historically attracted more attention.

Fig.1. Cognitive Architecture (ResearchGate.com)

PATTERNS OF CONDUCT

Different opinions can be found in the literature on what system can be considered a cognitive architecture. Cognitive architectures, on the other hand, must change through development and efficiently use knowledge to perform new tasks. Furthermore, he suggests toolkits and frameworks for building intelligent agents cannot themselves be considered cognitive architectures. (However, the authors of Pogamut, a framework for building intelligent agents, consider it a cognitive architecture. Another opinion on the matter is by Sun, who contrasts the engineering approach taken in the field of artificial intelligence with the scientific approach of cognitive architectures. According to Sun, psychologically-based cognitive architectures should facilitate the study of the human mind by modeling not only human behavior but also the underlying cognitive processes. Such models, unlike software engineering-oriented” cognitive” architectures, are explicit representations of the general human cognitive mechanisms, which are essential for understanding the mind.

In practice, the term ”cognitive architecture” is not as restrictive, as made evident by the representative surveys of the field. Most of the surveys define cognitive architectures as a blueprint for intelligence, or more specifically, a proposal about the mental representations and computational procedures that operate on these representations enabling a range of intelligent behaviors. However, given their prevalence in other areas of AI, deep learning methods will likely play some role in the cognitive architectures of the future.

To ensure both inclusiveness and consistency, cognitive architectures in this survey are selected based on the following criteria: self-evaluation as cognitive, robotic or agent architecture, existing implementation, and mechanisms for perception, attention, action selection, memory and learning. Furthermore, we considered the architectures with at least several peer-reviewed papers and practical applications beyond simple illustrative examples. For the most recent architectures still under development, some of these conditions were relaxed.

TAXONOMIES OF COGNITIVE ARCHITECTURE

Surveys of cognitive architectures propose various capabilities, properties, and evaluation criteria, which include recognition, decision making, perception, prediction, planning, acting, communication, learning, goal setting, adaptability, generality, autonomy, problem-solving, real-time operation, meta-learning, etc. While these criteria could be used for classification, many of them are too fine-grained to be applied to a generic architecture. A more general grouping of architectures is based on the type of representation and information processing they implement. Three major paradigms are currently recognized: symbolic, emergent, and hybrid.

Symbolic systems represent concepts using symbols that can be manipulated using a predefined instruction set. Such instructions can be implemented as if-then rules applied to the symbols representing the facts known about the world. Because it is a natural and intuitive representation of knowledge, symbolic manipulation remains very common. Although by design, symbolic systems excel at planning and reasoning, they are less able to deal with the flexibility and robustness that are required for dealing with a changing environment and for perceptual processing.

The emergent approach resolves the adaptability and learning issues by building massively parallel models, analogous to neural networks, where information flow is represented by the propagation of signals from the input nodes. However, the resulting system also loses its transparency, since knowledge is no longer a set of symbolic entities and instead is distributed throughout the network. For these reasons, logical inference in a traditional sense becomes problematic in emergent architectures.

Naturally, each paradigm has its strengths and weaknesses. For example, any symbolic architecture requires a lot of work to create an initial knowledge base, but once it is done the architecture is fully functional. On the other hand, emergent architectures are easier to design, but they must be trained in order to produce useful behavior. Furthermore, their existing knowledge may deteriorate with the subsequent learning of new behaviors.

CONCLUSION

Overall, in the literature on cognitive architecture far more importance is given to the cognitive, psychological, or philosophical aspects, while technical implementation details are often incomplete or missing. On visual processing, we list architectures that only briefly mention some perceptual capabilities but do not specify enough concrete facts for analysis. Likewise, justifications of using particular algorithms/representations along with their advantages and limitations are not always disclosed. Generally speaking, the lack of full technical detail compromises the reproducibility of the research.

Tutors India Uk Writers, this blog is completely free for educational purposes for the students exclusively who are pursuing the Masters and P.h.D. We are the #1 academic company, we help you in all kinds of dissertation topic writing and assistance based on your interest. Our academicians help you completely mail us to info@tutorsindia.com for further details

REFERENCES

Chiodino, E., Lieto, A., Perrone, F., & Pozzato, G. L. (2020). A goal-oriented framework for knowledge invention and creative problem solving in cognitive architectures. Proceedings of ECAI 2020, 2893-2894.
Guest, O., & Martin, A. E. (2020). How computational modeling can force theory building in psychological science.
Rich, P., Blokpoel, M., de Haan, R., & van Rooij, I. (2020). How intractability spans the cognitive and evolutionary levels of explanation. Topics in cognitive science, 12(4), 1382-1402.