Artificial Intelligence (AI) is a highly interdisciplinary field of research and development. It leans on many other fields, such as computer science, neuroscience, psychology, philosophy, mathematics and linguistics. Within the research and development of AI, as well as within the related research fields, there are countless disciplines and schools of thought with their own terminologies, approaches and methodologies. In order to be able to study architecture of AI systems, we do not want to position ourselves to any specific school of thought but try to establish the foundation for the architecture in field agnostic general terminology – starting from visiting the fundamentals of nature in its broadest sense and intelligence as a phenomena within it. This page presents introduction, foundations, approach and main research questions we identify for studying architecture of open artificial intelligence systems.
Fundamentals of Natural and Artificial Intelligence
Intelligence can be described as and ability of an entity to perceive information, retain it as knowledge and apply it to relevant behaviors in the environment and context of the entity. Intelligence is a natural phenomenon exhibited by biological organisms including bacteria, plants and animals – including humans. Among the forms of natural intelligence, human intelligence is the most widely studied area and stands out with high level cognition, motivation and self-awareness. Cognition enables us to learn, create concepts, understand and reason, while perceiving and interacting with our environment.
The nature constitutes the encapsulating environment for humans and human intelligence, including materials, physical objects and laws, as well as other organisms and humans within the same environment. Human life-cycle is governed by the nature and by human individuals behavior/interaction within it, including major events and phases in between: fertilization, fetal development, birth, childhood, adulthood, reproduction, death and disintegration. According to current knowledge and experience, humans have the ability to perceive their environment with senses (including at least: sight, hearing, taste, smell, hear, and touch) and actuate in it directly by using their body or indirectly by using tools and other artifacts made from natural resources (e.g. machines and robots of different kinds). The nature also sets physical and resource limits to all artifacts created by humans.
Humans are social and interact with their natural environment and other humans in their perimeter during their lifetime. The emergence of language for communication enabled humans to exchange knowledge with their peers, as well as transfer it to next generation. This enabled and initiated accumulation of knowledge, both during the lifetime of individual, and over generations of humans – a capability where humans excel in the nature. The accumulation of knowledge is accelerating catalyzed by series of human inventions, including writing & reading, books, book printing, systematic education of children, electronic communications, computers, the Internet and digitization of communication and contents.
An important observation of natural intelligence is also that intelligence is exhibited by individual organisms (a.k.a individuals/agents) and their co-operation. There is no known single organism holding collective intelligence in the nature of which an individual organisms would be a clone or subset of. Instead intelligence physically exists only during the lifetime of an individual organism and within the organism. Collective/social intelligence may be exhibited by co-operation of individual organisms (e.g a colonies of ants or human organisations/society), but it does not physically exist within a single organism. In nature, evolution over generations of individual organisms is modifying a genome, which can be seen as over-generation memory and instructions for prerequisites for birth of an individual intelligent organism.
Intelligence can also be exhibited by non-natural entities and is then referred to as artificial intelligence, where the intelligent entities are referred to as agents and are constructs of humans and human intelligence. In computer science Artificial Intelligence (AI) is also defined as the study and design of intelligent agents, where agent is a computing system that perceives its environment and takes actions to maximize its change of success at some goal. The performance of AI agents is often compared to human performance with the same task/goal.
Foundations for architecture framework of AI systems
The AI systems are bio-inspired systems, where design of a system is analogous to its biological counterpart. In case of human-inspired AI systems (our focus) the biological counterpart is human intelligence, cognition, brain, motivation, goals, behavior and performance. Accordingly, we take a bio-inspired approach to architecture of AI systems and we define the foundations of an architecture framework for AI systems in relation to their biological counterpart – humans:
- Artificial intelligence is constructed by humans with human intelligence. AI systems do not exhibit natural or human intelligence, but may mimic it as designed by their human creators. The human creators may exploit artificial intelligence to augment their own capability to design and realize new artificial intelligence systems but artificial intelligence is always a construct of humans and not part of the natural intelligence.
- Nature limits realization of artificial intelligence. As an human created artifact, realization of AI is limited by the nature as its encapsulating environment. Even though the inner “cyber world” of AI and actions within it are not limited by the laws of nature or physics (unless we want to simulate those), the realization of AI systems is limited in terms of availability of resources (e.g. materials for building supercomputers and energy for running those).
- Artificial intelligence may be both bio-inspired and “bio-integrated”. AI systems mimic biological intelligence but are also part of the same natural context and environment with natural intelligence – the nature. Even though AI systems behavior is artificially mimicking natural intelligence its perception and actuation can take place in the natural context and environment. Accordingly, AI systems perception and actuation can be seen as divided to two different worlds; the “cyber world”, when the operating context of the AI is purely artificial (e.g. within a computing system) and the “real world”, when the operating context of the AI includes sensing and actuating in natural environment (e.g. AI robotic systems).
- Artificial intelligence is exhibited by autonomous intelligent agents and their co-operation. Like in the nature, also AI is exhibited by individual artifacts/systems (intelligent agents) and by their co-operation with each other (intelligent multi agent system).
- Motivation and self-awareness of artificial intelligence is defined by humans. As an artifact AI does not posses any inner motivation or self-awareness unless the human creators of AI design such features for mimicking human intelligence.
- The life-cycle of populations in AI is governed by humans and limited by nature. The life-cycle of a human as biological organism is governed by the nature and human individuals interaction within it. As an artifact of human, the life cycle of AI is outlined by nature, but governed by the human creators and operators of AI. The life-cycle of intelligent agents differs from human life-cycle but can be compared: fertilization (idea/concept), fetal development (design), birth (development and deployment), childhood (training), adulthood (operational and learning), reproduction (versioning), death (undeployment) and disintegration (deletion). When comparing to natural human life-cycle, as artifacts, the intelligent agents have a possible additional non-natural life-cycle phase between death and disintegration: redeployment (which would be comparable to human resurrection).
- In order to exhibit intelligence an AI system has to be deployed, in an operational state and interact with its environment, consisting of nature and artifacts, other AI systems or humans, or a combinatoric system of these. The level of intelligence of an AI system is relative to its environment and is judged by humans observing its interactions and evaluating those. The evaluation may be based on comparing the outcomes to an optimal intelligence model (closed-world assumption of environment) or to value of the outcomes (open-world assumption on environment). For example, in a simple virtual game world, the optimal intelligence model results in actions that lead to winning the game or a draw. In the real word natural environments, the optimal intelligence model is usually not known or agreed on, and evaluation of the outcomes is based on value produced in the context of the observer.
- Human and artificial intelligence may share, exchange and transfer data, information and knowledge in shared representation. As long as data, information and knowledge is represented in a form understood by both artificial and human intelligence, these can be shared, exchanged and transferred between the two. It is noteworthy that if “knowledge” in AI systems is represented in a form not understandable to humans, this knowledge can not be exchanged with humans (e.g. neural network structures resulting from unsupervised deep learning). Similarly, the tacit knowledge of human intelligence is not exchangeable with artificial intelligence in absence of jointly understood representation for it – even among humans (e.g. riding a bike or all experiences during the working career).
- Both human and artificial intelligence can exhibit itself as collective intelligence. Human intelligence is exhibited by individual persons and collectively by co-operation of people in human organisations/society (e.g. cities, companies and nations). Similarly artificial intelligence may be exhibited by co-operation of intelligent artificial agents (systems) in multi-agent system (system of systems).
- Human intelligence and artificial intelligence are interconnected systems. Artificial intelligence is dependent on humans and human intelligence as creator and operator of it. The relationship between humans and AI is a human – technology relationship, which is limited and enabled by human – technology interaction. Technology serves the needs and desires of humans or human organisations and must comply with laws, social norms and other rules created by human organisations. The combination of human and artificial intelligence could be characterized as biological technological symbiosis on augmenting intelligence.
Approach to Study Architecture of AI Systems
There are many approaches and levels of detail to study of architecture of AI systems; One may for example be interested purely in the architecture of the technological implementation of an AI system in its different parts (e.g computing system architecture and hardware, software or communication architecture of the computing system). Whereas the architecture of technological implementation of AI systems is highly relevant for development and operation of such systems, a wider view to architecture and design of AI systems is needed in order to capture also the context, requirements and value of the AI systems in all possible environments for those.
The wider view that we take take to study of AI systems architecture is approaching AI systems in context of sociotechnical systems, where people-technology interaction can be considered from the viewpoint of individual person or from the viewpoint of human organization (e.g. workplace, company, nation and society). After all, the individual or organisational needs and desires of humans are the source of requirements and context for design, development and utilization for AI systems. The interconnection of human and artificial intelligence can be studied as a complex sociotechnical system of systems, where the behavior and social interaction of humans is interconnected with the behavior and interaction of intelligent agents in AI systems. This interconnection of human and artificial intelligence in context of sociotechnical systems is illustrated in the figure below.
Figure 1. The relationship of human and artificial intelligence.
The sociotechnical view to AI systems results into complex two dimensional interconnected systems of systems. The two main system dimensions are human social system and the related AI technology system. Furthermore, both of these dimensions have their own subsystems; AI agent can be considered as a subsystem in the AI technology systems dimension, as well as AI agents interacting with each other can be considered as a multi agent system – another type of subsystem in the AI dimension. Accordingly, each individual human can be considered as a subsystem in the human social system dimension, and a group of co-operating humans another type of subsystem. As illustrated in Figure 1, the interconnection between the two system dimensions is a human – technology interface and has a govern – serve type of relation between the dimensions; behavior and interactions between the intelligent agents and intelligent multi-agent systems are governed by behavior and social interaction of humans.
Even though the relationship between the human and AI systems dimension is characterized as govern – serve relationship, it can take many forms, from a strict rule based governance, to apprentice like co-operation based more on suggestions rather than commands. Technology has traditionally been built by humans as a tool to augment human capabilities of physical actuation (e.g. hand tools, engines and robots) or social interaction (e.g a letter, telephone, internet and social media) capabilities with unidirectional non-interactive interface. In other words, we do not (at least yet) expect a hammer shouting us hints on how to nail a specific type of nail to specific material. Neither do we expect to teach a hammer such knowledge that we may or may not have.
However, it is very natural for us to teach and guide our kids, co-workers and even dogs, and we also expect and enjoy interacting with them. What is the difference between a hammer and a dog? Well, a hammer is an artifact and a dog is a natural intelligent organism. If we want to augment our intelligence with AI technology we may also need to adapt our relation to technology in terms of governance and interaction with it. We may need to start teaching, guiding and interacting with AI technology more in same fashion that we interact with naturally intelligent organisms.
Human governance of AI systems covers the whole life cycle of AI systems from their design to their activation, operation and deactivation. An integral part of the governance and guidance between naturally intelligent organisms is different kinds of reward and punishment mechanisms. These mechanisms may need to be formalized and applied to governance and guidance of AI agents as well to enable autonomy, self-awareness and motivation of AI agents to perceive, reason, learn and actuate. In other words to enable AI agents to gather knowledge via experience.
Having presented our introduction via visiting fundamentals of human and artificial intelligence, we identified foundations and described our approach to study architecture of Opentech AI systems. Finally we describe a set of research questions that guide us in our way forward in study of Opentech AI architecture, and hopefully also appeal to other researchers and developers to get new knowledge on area of Opentech AI systems and their architecture:
- What are the distinguishing characteristics and features of Opentech AI systems and how those can be captured and considered during the system life cycle?
- What are the key characteristics and components of an AI agent and what architectures, methods and tools exist for rapidly re-building AI agents from existing sub-components?
- Is there a basic organisation of an AI system that is application independent and what are the main views and functional components of an AI system?
- Could such an organisation be defined as architecture framework for describing architectures of AI systems?
- Could such an organisation be used for systematically measuring progress in various AI systems R&D?
- How to increase modularity, interoperability, portability and evolvability in AI systems development in order to enable reuse and speed-up development of AI systems?
- What is the full (allopoietic and autopoietic) life-cycle and context of an AI system and how it relates to existing development and operation practices of software intensive computing systems?
- What kind of stakeholders and roles are generally related to AI systems during their full life-cycle?
- Can knowledge of AI systems be accumulated during the full life-cycle and what are the requirements for efficient knowledge accumulation of AI systems?
These are the questions we are pondering in our minds and we will post some initial results addressing these research questions as the work proceeds.