The Principles of AI and Causal Intelligent Machines
We update McCarthy's classical definition of Artificial Intelligence (AI) as "the science and engineering of making intelligent causal machines”.
True AI and Machine Learning is to automatically and autonomously identify, understand, explain, infer, discover or predict possible causal variables, mechanisms, patterns, rules and laws, systems and networks from the universe of data to interact with the world in the most effective, efficient, rational and sustainable ways.
The Fundamental Principles, Assumptions and Axioms
Principle 1: Science and Technology and Human Practice are all about discovering and applying causal relationships, their processes, mechanisms and effects
Principle 2: Intelligence consists in simulating and modeling, understanding and effective interacting with the world and reality, its causality and causation, variables and interactions, causative patterns, rules and laws of reality
Principle 3: Artificial Intelligence consists in Causal Machine Intelligence and Learning
Principle 4: Machine Intelligence postulates Causal Intelligence
Principle 5: Machine Learning postulates Causal Learning
Principle 6: No Causal Intelligence, no Machine Intelligence
Principle 7: No Causal Learning, no Machine Learning
Principle 8: Non-Causal Machine Intelligence and Learning are Non-Intelligent AI
Principle 9: Non-Causal Machine Learning is not Real Machine Learning
Principle 10: Real/Causal AI embraces all the ethical principles for responsible and trustworthy AI (as OECD AI Principles )
General Observation
The prevailing AI and ML is not real and true machine intelligence but imitative human intelligence applications; for it does not take causality and causation into account, and instead is concerned with prediction based on statistical associations.
This observation relates to all the large-scale language machines, LLMs, from Google's LaMDA to Microsoft/OpenAI' GPT/ChatGPT.?
As a result, all LLMs, as LaMDA, GGT-3/4, ChatGPT, OPT-175B, Megatron-Turing (MT-NLG), PaLM, Gato AI,?the transformer machine learning models,?are false positives.
It is when a test result incorrectly indicates the presence of a state or condition (such as a disease or pregnancy or intelligence, when it is not present), while a false negative is the opposite error, where the test result incorrectly indicates the absence of a condition when it is actually present. Or, rejecting the null hypothesis or not.
Transformers , taking advantage of parallel computing hardware and graphics processing units (GPU) in training and inference. like other neural networks, are ONLY statistical models that capture regularities in data in clever and complicated ways, having no language understanding or any intelligence.
False and Fake AI/ML (FFAI/ML), which is the mainstream AI, going as Big Data Analytics, Narrow/Weak AI, ML, DL or ANNs, LLMs, NLP, Computer Vision, Machine Perception.
As such, the FFAI/ML is about faking/mimicking/replicating/simulating the human body/brains/mind/intelligence/cognition/behavior.
The FFAA is by its very definition is harmful for humanity; for it is designed not to augment and enhance, but to replace humans, our jobs, works and positions, as less creative and innovative, but more efficient and productive.
Commentary
The biggest trend in AI and ML and DL
The biggest trend in AI and ML is building Causal Machine Intelligence and Learning. The idea is to integrate causality, learning causal patterns, correlations and associations, rules and laws and effects from data.
Machine learning based on statistics and probabilities, correlational pattern recognition, like LMMs, as ChatGPT or LaMDA, is insufficient for meaningful predictions and reliable decision-making.
New approaches to machine intelligence and learning (MIL) based on principles of causal reasoning, modeling and analysis provide a promising path forward to Real/True/Generalized AI.?
The idea of?causal/physical/real AI ?was first generated as part the Encyclopedic AI project (A. Sh. Abdullaev, Preprint Knowledge Base of Encyclopedic Artificial Intelligence, Moscow, 1989, Academy of Sciences of the USSR).
The key assumption is “Machines' lack of understanding of causal relations is perhaps the biggest roadblock to giving them human-level intelligence.” [The Book of Why: The New Science of Cause and Effect, 2018]
In reality, it is humans' lack of understanding of causal relationships as interactions or productive correlations and associations is the biggest roadblock to giving machines human-level intelligence.
Causation postulates Correlation, and Correlation implies Causation
Correlation?means there is a statistical association between variables.?Causation?means a real association between variables when a change in one variable causes a change in another variable, and VICE VERSA.
Correlation and causation are two entangled things at all levels, from the quantum level to the cosmological level of the whole universe: “correlation implies causation”, while "causation postulates correlation".
First of all, there is one real relationships in the world and reality, which is causation or cause and effect, all the the rest, as resemblance, contiguity, or wholeness, have the less reality, if any.
Any mature science and good researchers in any domains ?are implicitly guided by the ten fundamental features of real causation:
"The cause precedes the effect (time order). The cause co-occurs with the unaffected entity in space and time (co-occurrence). Causes and their effects are the result of a web of causation (preceding causation). The intensity, frequency, and duration of the cause are adequate and the susceptible entity can exhibit the type and magnitude of the effect (sufficiency). The cause effectively interacts with the entity in a way that induces the effect (interaction). And, the entity is changed by the interactions with the cause (alteration)".
Besides, causality has all the major attributes of relationship: symmetry, transitivity and reflexivity.?
Still, many dogmatically presume that "correlation doesn't imply causation" and that causality is the asymmetrical relations of cause and effect, the cause (independent variable) must precede the effect (dependent variable) in time, as?the treatment/intervention 'causes' the outcome in the RCTs (A/B) experiments.
Causation is defectively defined as?the relationships between independent, control, and dependent variables, as?a linear, asymmetrical relationship between two variables X and Y such that changes in X lead to changes in Y, and the key difference between association and causation lies in the confounding.
The causal metamodel is confused as a three-level abstraction, the ladder of causation, as
associations/probabilities/correlations/regularities/patterns (seeing/observing);
causality/intervention (doing);
counterfactuals/but-for causation/sine qua non?causation/physical laws (thought experiment/imagination/reasoning). [ Pearl, J.?Causality: Models, Reasoning, and Inference. Vol. 64 (Cambridge University Press, 2000)].
This all corrupts the whole science and engineering of causality and causation, the real nature of causal processes and mechanisms, causal analysis (regularity, probabilistic, counterfactual, and manipulative), reasoning and inferences and models, impact, efficacy and effectiveness, measures and effects.
The linearity of causality brings to the poor conception of causal graph or causal network (also known as path diagrams,?causal Bayesian networks ) as a topologically ordered directed acyclic graph, DAG, with no directed cycles, where one event (X) is to cause another if it raises the probability of the other (Y).
In statistics, econometrics, epidemiology, genetics and related disciplines, causal graphs as DAGs are?probabilistic graphical models encoding assumptions about the data-generating process. Such?statistics revolves around the analysis of relationships/associations/correlations among multiple variables.
In reality, causal structure or causal system is the network graph of causal interrelationships [associations/correlations/interactions] among or between a set of causal variables, and causal discovery algorithms are the problem of learning the causal structure, patterns and correlations, rules and laws, from data, observational and experimental, simulation or theoretical, real-world or synthetic, structured or unstructured, digital or analogue.?Real machine statistics revolves around the analysis of causal relationships among multiple variables/nominal, ordinal, interval, ratio, numeral data.
Causal MIL vs. Causal AI
Gartner has included causal AI in its 2022?Hype Cycle ?report, citing it as one of five critical technologies in accelerated AI automation.
领英推荐
Causal AI (CAI) ?is defined as a statistical?AI (SAI)?system that can explain [linear]?statistical causality , causation, or cause and effect, and the CAI technology is to help explain decision making and the causes for a decision.
Among the developers of stochastic causal AI software are Microsoft , causaLens, Xplain Data,?Geminos?and Qualcomm.
Causal AI ?autonomously searches for causes in data, while also boosting experimentation and human intuition.?Causal AI discovers the underlying causal relationships in data, whereas machine learning just analyses correlations.?
Causal knowledge can be acquired in three mutually complementary ways:
through?experimentation;?via?human expertise?and intuitions; with?causal discovery?algorithms
"Causal AI frameworks and algorithms support decision making tasks like estimating the impact of interventions, counterfactual reasoning and repurposing previously gained knowledge on other domains".?
Decision-making AI
Causal AI doesn’t just predict the future, it shapes it.
Explainable AI
Put the “cause” in “because” with next-generation explainable AI.
Adaptable AI
Causal AI continuously adapts to real-world dynamics.?
Human-centric AI
Human-plus-Causal AI partnership allows organizations to harness the benefits of AI.?
Imaginative AI
Causal AI can explore hypothetical worlds, uncovering insights that explain why events happened.
Fair AI
AI has a bias problem and Causal AI is the solution.
AI for small data
70% of organizations are shifting their focus from big to small data — Causal AI can help.?
Trustworthy AI
Trust is the most important but often-overlooked ingredient in successful AI adoption.?
How are Causal AI models different from Bayesian networks?
The two types of models have some superficial similarities, but they also have significant differences. Bayesian networks (BNs) simply describe patterns of correlations between variables. Causal AI models capture the underlying processes that drive those statistical relationships.
This paradigm shift makes Causal AI models more flexible, versatile, and powerful than Bayesian networks.
In fact, a?Bayesian network ?(a?Bayes network,?Bayes net,?belief network, or?decision network) is a probabilistic?graphical model?that represents a set of variables and their?conditional dependencies?via a?directed acyclic graph?(DAG), where efficient algorithms can perform?linear causal inference and learning.
What we really need is realistic generalizations of Bayesian networks, in which not only?probabilistic inference or decision making?problems (following the?optimum of objective, loss or reward functions, as the maximum expected utility?criterion) can be modeled and solved.?
Causal Machine Intelligence and Learning (CMIL)
CMIL/AI Engineering: Reality > Causality = Interaction > Real Science and Technology > Artificial Intelligence and Machine Learning > Statistical AI/ML/DL > Causal AI > Interactive AI = Real Machine Intelligence and Learning > Trans-AI
Unlike a narrow and weak AI and causal AI, CMIL/Real AI is a generalized intelligence, integrating symbolic AI, statistical AI/ML/DL and causal AI with human intelligence.
RAI is driven by its Global Intelligence Engine, the World Hypergraph: Global Causal Graph Network: Interactive Bayesian Graph Networks, governed by the?general?chain rule?(the?general product rule) ,?where all the world's variables (causal factors) are interacting (conditionally dependent of one another) with each other.
It underlies the possible architectures of artificial neural networks, as the most advanced deep neural networks algorithms.
RAI has knowledge, competence, skill or?literacy ?in ANY field of knowledge or practice:?science literacy; technology literacy; computer literacy;?statistical or data literacy;?critical literacy;?media literacy;?ecological literacy,?disaster literacy;?health literacy;?social literacy;?quantitative literacy (numeracy),?visual literacy, e.g. body language, pictures, maps, and video), etc.
The CMIL/RAI Enginery includes:
Conclusion
As such, there are two classes of AI:
Real and True AI as Causal Machine Intelligence and Learning.
False and Fake AI/ML (FFAI/ML), which is the mainstream AI, going as Big Data Analytics, Narrow/Weak AI, ML, DL or ANNs, LLMs, NLP, Computer Vision, Machine Perception.
The FFAA about mimicking/replicating/simulating the human body/brains/mind/intelligence/cognition/behavior
The FFAA is by its very definition is harmful for humanity; for it is designed not to augment and enhance, but to replace humans, our jobs, works and positions, as less creative and innovative, but more efficient and productive.
Supplement: OECD AI Principles overview
An AI system is a machine-based system that is capable of influencing the environment by producing an output (predictions, recommendations or decisions) for a given set of objectives.
It uses machine and/or human-based data and inputs to (i) perceive real and/or virtual environments; (ii) abstract these perceptions into models through analysis in an automated manner (e.g., with machine learning), or manually; and (iii) use model inference to formulate options for outcomes.
AI systems are designed to operate with varying levels of autonomy.
AI system lifecycle phases involve: i) ‘design, data and models’; which is a context-dependent sequence encompassing planning and design, data collection and processing, as well as model building; ii) ‘verification and validation’; iii) ‘deployment’; and iv) ‘operation and monitoring’. These phases often take place in an iterative manner and are not necessarily sequential. The decision to retire an AI system from operation may occur at any point during the operation and monitoring phase.
AI knowledge refers to the skills and resources, such as data, code, algorithms, models, research, know-how, training programmes, governance, processes and best practices, required to understand and participate in the AI system lifecycle.