The home of innovation in

At its core AI is a process of reducing uncertainty in situations. AI algorithms and representations must be able to intelligently handle complex real world-situations. They must also be able to reason with significant imprecision, vagueness, multiple truths and a ubiquitous breadth of varied uncertainties to be able to solve problems or predict consequent world states. Such an AI is required in cognitively complex domains (e.g. human behaviour, forecasting, logistics, prediction, etc). Recognition based AI is not sufficient; what is needed is AI with adaptive comprehension and understanding of all the realms of uncertainty.

In contrast to many of today’s AI technologies, which sweep polyform uncertainty ‘under the carpet’, we promote the use all the varieties and sources of uncertainty in the cognitive process to emulate the subtlety of acumen exhibited by human intelligence. Our platform has a foundation of analogy reasoning and learning which exploits advanced rationalising technologies to enable graceful cognition.

Soft computing is fundamental to cognitive AI as it is able to deal with vague forms of knowledge, situations, reasoning and learning, and manage simultaneous multiple contexts/motives in an adaptive way.

Cognition is acquiring knowledge, understanding and truth through thought, experience, the senses and also partially through recognition. It is the combination of knowledge and reasoning to predict new implicit unobserved insights (data, facts, information, knowledge, wisdom, etc). This involves comprehending and understanding the world using inherently symbolic representations and processes.

In contrast, recognition is the acknowledgment of something's existence, validity, or legality. It is essentially labelling an ‘item’ with a ‘symbol’ through superficial pattern recognition. This can be either a sub-symbolic or a non-symbolic process, it does not need the impenetrable contortions inherent within neural computing.

Explanation, justification and transparency of insights are by their very nature fundamental methods for cognitive and language-based processes.

In contrast to many of today’s machine learning algorithms which are focused on just recognition, Inferz is focused on human-like cognition capable of reasoning under uncertainty with multiple world views. Inferz’s technology is capable of both soft pattern-based recognition/perception and more powerful language/knowledge-based inference i.e. representational cognition.

Inferz’s algorithms/representations are fully justifiable (it is clear exactly how and why each conclusion is reached). Both procedural and causal explanations are available. This is of particular importance in fields such as medicine, engineering and law, where auditability is required for practical or regulatory reasons.

Each cognitive decision or situation state clarification is justified by:

• facts and certainties that are associated with the current contexts

• real-time working model information

• domain models (ontology based)

• machine learned experience held in the knowledge bases

All of these sources are used to draw conclusions, explanations, justifications, levels of certainty, causalities and clarifications via machine reason-based auditing inference.

​

Connectionist AI has only reached the simplest stages of recognition, equivalent to the neural power and awareness of a slug. It will always be limited to recognition tasks only. It is not self-aware, because its processing is simplistic, obscure and opaque with no representational decision method, which is essential for justification. As such it is impossible for connectionist AI to explain or clarify itself. Even more crucially it has no human-related model of the world to articulate ‘why’ or ‘causation’ to a human. It has no meta-knowledge, understanding or meta-cognition.

Polyform learning is a holistic approach to machine learning which is able to utilise all forms of uncertainty and complexity.

Fundamental to its success is the soft computing principle of combining several treatments of uncertainty as an intrinsic part of learning and synthesis.

There is a need to incorporate the following approaches to handle the complexity of human-like learning: fuzziness/approximation, probability/likelihood, belief/truth, entropy/information-disorder, etc. Machine learning engines typically have only one method of handling uncertainty or none at all: in either case this results in naïve AI.

Along with the handling of multiple forms of uncertainty, machine learning engines must also be able to: test the validity of the resultant synthesised knowledge through understanding ontological sense, explain the source of information from data, explain the derivation of that knowledge and explain the derivation of predictions from that knowledge.

Machine learning can be a powerful knowledge synthesis technology, but without machine reasoning the AI system will only execute/invoke previously synthesised patterns. They will only perform recognition not true cognition. They will lack the ability to infer or create unseen implicit situation consequences.

Such reasoning should be able to fully discern the consequences of subtle uncertainty (i.e. fuzziness, probability, bias, belief, truth, etc) in the world. AI limited to only machine learning will merely achieve ‘idiot savant’ status in limited domains and simple consultation paradigms. Such AI will only appear intelligent, like a Victorian automaton – fascinating but fake.

Instead AI needs to be able to solve deep complex problems/situations, it needs to be dynamically adaptive using a variety of reasoning techniques (i.e. inference: deductive, inductive, abductive, etc; also theorem provers, constraint satisfiers, linguistic reasoners, etc) and to be aware how/when to invoke each technique. It needs to recognise entities/situations but it also needs to understand, be aware and be astute to attain true cognitive capabilities.

All this requires explicit symbolic reasoning - logical, non-logical and knowledgeable. The reasoning process should be fast and scalable, exploiting modern computation power, parallel processing and coherence-based computing.

Polyform cognitive processing is a holistic approach to both machine learning and machine reasoning (ML/MR). Fundamental to cognitive ML/MR’s success are the principles encompassed in soft computing, i.e. utilising multiple treatments of uncertainty as an intrinsic part of representing, learning, discovering and subsequent predicting. Polyform AI combines and incorporates the following technologies to handle the complexity of human-like ML/MR: fuzziness/approximation, probability/likelihood, belief/truth, randomness/entropy and others.

Polyform AI recognises the hyper-dimensional nature of ML/MR, not just the multiple attributional nature of the problem space and its curvature, but also the associations in the space - degree of set membership, the probability of set assignment, the degree of belief in the set information; the hyper-associations/relations/sequences present in the space; the extra-dimensionality facets of time, events and the like; and the meta-dimensionality properties of the solution search space such as context, conjecture, hypotheses, assumption, etc.

Polyform AI also recognises a myriad of forms and types of insight, wisdom, knowledge and information.

In addition to explainability and justification, a real-world AI requires:

​

• Transparency – not only making explicit the derivation of the facts in the world model/situation but also defining the different forms and magnitudes of uncertainties in those facts, decisions and observations. Such uncertainties include – bias, approximation, likelihood, and others.

​​

• Context of truth, a maintenance of truth and the degree of assumptions in the truth.

Transparency should involve all the artefacts of reasoning (and learning) from explicit content, data and information, derived implicit data/information/facts, and the invoked knowledge/wisdom/insights and how they are utilised.

These artefacts are multifarious and have context and uncertainties accompanying their use in the reasoning and broader cognitive process.

Vague rationale, as embodied in approximate (and probabilistic) reasoning, results in solutions that avoid the false beliefs and oversimplifications inevitable in Boolean/Bayesian hard logic systems.

Progressive qualification and endorsement of decisions under uncertain or ambiguous facts/knowledge is how humans deal with the dynamic shifting nature of reality. Also the real world is not precise or cut-and-dry: it is a complex world that naïve neural recognition technologies oversimplify, leading to brittle problem solving.

Cognition is by its nature able to operate with vague inference when dealing with intricate, imperfect and imprecise continuums of understanding that are all-pervasive in the world and how it is represented.

AI must be able to reason and learn with progressive gracefulness as the world gets more imprecise, indiscrete and indeterminate.

The world and its problems/opportunities are easier to understand if they can be conceived on the basis of an underlying model. Such a model should be representational and have sufficient erudition for the performance of cognitive tasks and be complementary to language-based elucidation. This model should be explicit but conceptualised from first principles that can encompass both competing ideas and many aspects of uncertainty – whether they are intuitive or counterintuitive. This model should be a repository of ontological forms of knowledge. Based on an intrinsic core ontology, it is required that machine classification of general/domain centric content should continuously refine/discover new ‘extrinsic’ ontological items, associations and uncertainty characterisations.

Soft machine learning is an ideal method for generating the ontology using fuzzy clustering with probabilistic and entropy-based metrics. These metrics should guide the characterisation of the linguistic entities and soft semantic networks. Polyform fuzzy clustering and comparative soft truth maintenance is ideal to enable the autodidactic learning and discovery of ontological knowledge.

The ontology will be the underlying world model for focusing general/domain/language-based reasoning using conditional knowledge that has been garnered by sub-symbolic and symbolic soft machine learning.

Cognitive AI requires a move from the simplistic in-vogue statistical based Natural Language Processing (NLP) to embrace Natural Language Understanding (NLU). Using statistical NLP we lose all context and meaning resident in the textural content as it is converted into just data about the words. We lose all semantic meaning and the nuance inherent in the imprecision/vagueness of words or their settings. NLU based on statistical correlation loses the ability to reason deeply about language usage and world insights expressed in the ambiguity of the language.

Linguistic intelligence based around language models, world models and language-based reasoning is an adjunct to domain-based problem solving. It uses a parallel approach of understanding and reasoning about concepts framed in words. It aims at enhancing the AI’s ability to think constructively about what is being communicated and why: to think, reason and solve problems in different ways. This is very unlike simple vocabulary recognition. It leverages the intelligences held in the linguistic propositions leading to arguments and reasons by which to enhance discovery.

Language is much more than the shallow understanding of words, their meaning, their usage and their semantics. Language usage encodes not just description/enquiries but importantly the uncertainty and wisdom of a situation through the choice of words and their articulation. Crucially, the expression of vagueness/imprecision with regards to the meaning (semantics), our level of knowing (epistemic), the inherent fuzziness and likeliness due to genuine indeterminacy in the situation (ontic) and our ability to describe a situation (nomic) are a valuable insight for reasoning about the world. Currently this meta-knowledge is ignored or ill-used by non-symbolic AI. In Cognitive AI, the use of vagueness is fundamental to understanding our imperfect real world and places caveats on our levels of precision to reason about it. It is the way the human mind works. It is also fundamental to the AI’s ability to communicate such reservations in situations or solutions. Ethical AI is a direct consequence of the recognition of uncertainty and bias.

Non-symbolic statistical word matching loses all this necessary capability. What we need is an ontology-driven linguistic reasoning technology which can exploit and use the uncertainty insights bound and encoded in the written word. By adopting deep linguistic learning and reasoning, embodying the power of fuzzy logics, we can emulate common-sense and maintain complete transparency of the resultant cognitive knowledge and acumen.