Latest Research

Our proposed solution to addressing global challenges and future challenges, our vision of future intelligence, latest research reports and Research Proposals.

Global Challenges

cited from Imperial College, London

• Climate Change and Environment
• Global Health
• Energy Crisis
• Security (data, cyber, war, terrorism)
• Data Science
• Molecular Science and Engineering
• Infection

Complex Systems

Complex Systems are pervasive and the most exciting and intriguing fields and challenges for future intelligence. Certainly they must be decomposed into smaller manageable subproblems to be tractable.

• healthcare
• robot football
• wargaming
• world economy
• stock market
• internet
• atmosphere
• and more…

Future Challenges

• Complex Systems
• short-termism
• exploration of the unknown, such as a new medicine or interstellar travel
• and more…

Unknown Threats

• asteroid impact
• global infection
• alien civilisation invasion
• and more…

Intelligence

The following characteristics are indispensable for both intelligence and future AI:

• Understanding: one of the pivots
• Learning: based on understanding, for generalisation
• Reasoning: the basis
• Intention: for the benefits of the subject
• Optimal individualised decision: the intervention
• Explainability: based on understanding, related to abstraction
• Persuasion: based on understanding, for better decision
• Reflection and remediation: punish/reinforce/modify/change based on assessment of previous decision
• Adaptation and evolution: for better intelligence
• perception, knowledge representation, communication, interactions, similarity
• High Intelligence:
  • Abstraction and inverse-abstraction
  • relevant cause-and-effect analysis
  • Decomposition of complex problem into probabilistic optimal substructure and probabilistic overlapping subproblems, then parallelising them
  • Automatic translation between different knowledge representations
  • creativity, innovation, analogy, inspiration, association, intuition, insights, etc.
• Other aspects…

Reasons present AI cannot match human domain expert intelligence in critical areas

critical areas include healthcare, self-driving, aerospace, finance, etc. where safety and accuracy are crucial

• still elementary intelligence: probability calculation and repetition, therefore unable to generalise well
• insufficient data or small dataset for training
• lack of high intelligence like abstraction, inverse-abstraction, analogy, etc.
• missing key background partial-knowledge beyond pure data distributions
• lack of data for relevant cause-and-effect analysis
• lack of automatic translation between different knowledge representations

Illusions of present AI

• AlphaGo could defeat human world champion because of the enormous computing power behind AlphaGo, but it could be 100% easily defeated by specially designed strategies of human high intelligence, therefore AlphaGo is still elementary intelligence
• LLMs could partially understand/reason due to massive training datasets which are nearly all the knowledge of human civilisation. They are unreliable and can’t generalise well, transformers are designed for processing inconstant-length sequential data and (long) dependencies, like (natural/computer) languages. However they are clumsy at processing non-sequential data and could play an important supporting role in future intelligence

Solutions

[Future Complex summary]

Proposed solution to addressing global challenges and future challenges

Human civilisation is facing Global Challenges [2] such as Global Health [4], future challenges such as Complex Systems [3], unknown threats such as global infection. To address these challenges, Future Intelligence (FI) equipped with Super Auxiliary Abilities (SAAs) is required, furthermore, a distributed system of decentralised, heterogeneous and specialised Units which are consisted of combinations of FI/SAAs is necessary. Synergies between Units, such as complementarity, cooperation, collaboration, coordination, management, communication, explanation, negotiation, making group decisions, could exponentially augment the capacities of such a system as a whole. We therefore envisage such a system as Future-Complex whose overall capacities could reach far beyond existing human intelligence and abilities, and is designed to assist human civilisation to addressing those challenges existing intelligence and technologies unable to handle. Unprecedented Future Intelligence will be at the centre of Future-Complex and will be a combination of Biological Intelligence [5], future Emotional AI and future Rational AI. ProfoundAI (PAI) [1] is our vision of future Rational AI.SAAswill be less-intelligent yet significant and innovative technologies, such as wireless devices, biomarker analyser, CT (Computerised Tomography), physical therapy (like ultrasound, microwave, infrared), clairvoyance, clairaudience, etc. Units could be either embodied existence, such as embodied AI [6], or disembodied existence. Their communications and interactions could be resolved by existing technologies such as IoT, blockchain, Large Language Models, etc., provided existing issues of which could be solved. Further studies of Future-Complex could be directed by Complex Systems theories [3], such as collective behaviour, fractals, emergence, game theory (collective intelligence), dynamic systems, Cybernetics etc., and certainly joint efforts of multi-disciplinary research. Highlighted benefits of Future-Complex to addressing Global Health challenge include telemedicine; understanding of the disease and intention of patients; tracing causes of diseases and making prognosis; automatically measuring biomarkers and early warning of risks; providing emergency assistance; finding optimal individualised treatment; explanation of diseases and treatments based on evidence; optimal multi-morbidity treatment consultation; robot assisted surgery; simulation and prevention of global infection; effective global coordination of pandemic control, etc.

[Future Intelligence and Profound AI summary]

Our vision of future intelligence

Present stage AI like Large Language Models (LLMs) and AlphaGo are still elementary intelligence [12] and people are bewildered by the enormous computing power and massive training datasets behind them, they could neither generalise well nor match human domain expert intelligence in critical areas like healthcare/self-driving because of several reasons: real-world issues are complex systems which are computationally intractable and impossible to accurately simulate; absence of high intelligence [12], including abstraction [5] (like rules, concepts, principle, strategies), inverse-abstraction (applying rules, concepts etc. for prediction), similarity, analogy etc.; insufficient data or small dataset for training; lacking key background partial-knowledge; missing relevant information for cause-and-effect analysis; lacking automatic translations between different knowledge representations. Profound AI (PAI) is our vision of future Rational AI that aims to remedy these deficiencies, prominent features of which include: perception of previously undetectable; similarity reasoning; understanding; finding optimal individualised (sequential) decisions under uncertainties; explainability; intervention; control of peripherals; persuasion; communication; exchange; interactions; assessment of outcome; reflection and remediation; simulation of complex systems; evolution and adaptation through new data/observations/experiments/evidence. It also has high intelligence including: abstraction and inverse-abstraction; decomposition of sophisticated problem into smaller manageable subproblems via abstraction and combinatorics, probabilisticoptimal substructure and probabilisticoverlapping subproblems with adaptive partial-optimal decisions, then parallelising them; adaptive behavior/strategy by observing environment/opponent/teammate and predicting their behavior/strategy; different knowledge representations at different hierarchical levels and automatic translations between them; explain/persuade in the object’s personal/domain/subgroup knowledge representations (language/experience/sense/knowledge-base/knowledge-structure); cross-domain generalisation; chain of cause-and-effect analysis; analogy, association, inspiration, intuition, insights; genius intelligence like creativity and innovation; super human intelligence like super memory, knowledge of whole human civilisation, instant understanding/learning/reasoning. Furthermore, appropriate (reasoning) architectures/algorithms are necessary depending on the specific datasets/tasks/domains and variable relationships, such as linear, nonlinear, (uni-/bi-) directional, time-series (unidirectional), sequential (bidirectional), (positive/negative/neutral) correlation, inconstant-length, (long) dependency, regional, tree-like, graphical, partially-observable, deterministic/non-deterministic/probabilistic, confounders/covariates, hidden-factors, different dimensional spaces (from zero to high dimensions), rules, logic, step change (threshold), continuous, discrete; these could be resolved by MLP/CNN/GNN/XGBoost/Transformer etc. or completely new architectures/algorithms. Future Intelligence (FI) will be high intelligence, genius intelligence, collective intelligence, possibly super natural intelligence and unimaginable intelligence; we envisage that FI will be consisted of Biological Intelligence (BI) [9,11], future Emotional AI (EAI) and future Rational AI (RAI), each sub-component has both advantages and disadvantages, and complements each other, synergy of these sub-components will take Future Intelligence to a new level far beyond existing human intelligence and present stage AI. We argue that PAI will be future RAI. The BAI-PAI project family are preliminary implementations of PAI, which are consisted of BAI-OPTIMA [14], TAIR-CDSS [10], BAI-RA [13] and several Research Proposals [7,8].

[BAI-OPTIMA project summary]

BAI-OPTIMA is a preliminary implementation of PAI 

It is still difficult for present stage AI to match human domain expert intelligence in critical areas like healthcare/self-driving/financial-sectors due to several reasons [2] which Profound AI (PAI) [2] aims to remedy. To this end, abstractions and inverse-abstractions (applying abstractions like rules) are necessary to reason beyond limited data distributions, which enable small dataset training and (limited) generalisation. BAI-OPTIMA (BenAI common OPTimal Individualised decision MAking architecture and algorithms for critical areas) has the ambition to be a common decision making foundational framework for critical areas, provided variables relationships identified, appropriate reasoning architecture found, generalisation achieved and corresponding Domain Knowledge imported. The GID model is a sophisticated ensemble model consisted of an untrainable abstraction module (like medical rules), a trainable abstraction module (representing Domain Expert Experiences), a prediction module, an abstraction-enforcement module (enforcing expert’s experiences), a utility function. By introducing Domain Knowledge into the model, it supplements insufficiency of small training datasets. For deterministic decision making, it could find the triggering thresholds. For non-deterministic decision making, optimal (individualised) decision is based on the matching rate of the decision and corresponding predicted outcome. At the moment, it could process structured data like numerical, categorical, date-time, etc. and will be extended in the near future.

[Major innovations of BAI-OPTIMA]

please refer to BAI-OPTIMA project for [details]

[TAIR-CDSS summary]

Decision making in healthcare domain based on BAI-OPTIMA project

TAIR-CDSS (Thick AI Reasoner for Clinical Decision Support System) is based on the BAI-OPTIMA [4] framework and an application of decision-making in healthcare domain demonstrated by asthma treatment. By incorporating Domain Knowledge into asthma treatment, TAIR-CDSS knows the symptoms and causes of asthma, pathology and pharmacology, medical rules, therefore logically understands [1,3] asthma. TAIR-CDSS is consisted of 4 layers 1) diagnose 2) computing the matching rate of individualised treatment by simulating a specific medical specialist 3) make prognosis 4) compute optimal individualised treatment out of the matching rate of the treatment and corresponding prognosis. It could explain medical reasons of the treatment from guideline and medical evidence, induce experiences of asthma specialist into rules, find subgroup of asthma treatment, find similar historical cases; evolve through Evidence-Based Medicine (EBM). The first phase of the project is near finishing. 

Please refer to TAIR-CDSS project for [details]

Research Proposals

Step by step towards future intelligence

• [RP game theory]
• [RP abstraction reasoning]
• [RP similarity reasoning]
• [RP intuition reasoning]
• [RP inspiration reasoning]
• [RP analogy reasoning]
• [RP association reasoning]
• [RP ambiguity reasoning]
• [RP auto translation]
• [RP uncertainties reasoning]
• [RP synergy]
• [RP theory]

Deficiencies of human intelligence that could be complemented

• lack of enormous computing capabilities. possible solution: AI
• lack of highly effective collaborations. possible solution: collective intelligence
• lack of super assistive abilities. possible solution: SAA like CT and wireless communication

The role LLMs will play in future intelligence

Processing sequential data like auto translation between different knowledge representations provided generalisation could be resolved via abstraction, inverse-abstraction and analogy

Deficiencies of evolution of human intelligence and human civilisation 

Avoided the worse case but cannot guarantee the best case