Unifying Generative AI, Machine Learning, and Rules-based Systems

In an older post, I mentioned that there were two types of artificial intelligence: rules-based and patterns-based.  I also mentioned that you should use the right tool for the job.  This was before the ChatGPT fad started.  Now it appears that there are three general categories of or approaches to artificial intelligence which are:

  • Generative AI: Generative AI is a subset of artificial intelligence and a type of machine learning that produces new content from existing content.  Generative AI uses large language models (LLMs) and transformers to take natural language and synthesize new information. ChatGPT is an example of generative AI.
  • Patterns-based (machine learning): Machine learning is a subset of artificial intelligence that uses different types of algorithms that leverage statistics and probabilities to analyze data, learn from that data, and imitate the ways humans learn and then make predictions and informed decisions based on the data. Because machine learning is based on probabilities, it can never be guaranteed to be correct.
  • Logic and rules-based approach (good old fashioned expert systems or symbolic AI): Logic and rules-based artificial intelligence uses machine-readable representations of information perhaps in the form of a knowledge graph and logic engines that can read and process that information to mimic human thinking. This approach provides certainty in terms of the result.
What most people don't seem to understand is that all three of these types of artificial intelligence are tools that are available and that the RIGHT TOOL SHOULD BE USED FOR THE JOB.  None of these approaches are "best" in all use cases.  Rather, you should understand your use case and understand the capabilities of the tool and then match the tool with the use case.

Another approach is to unify the approaches, creating a hybrid, and then you can leverage the best characteristics of each approach to meet the needs of your use case. The article Unifying Large Language Models and Knowledge Graphs: A Roadmap, provides this very nice comparison of the PROS and CONS of knowledge graphs (KGs) and large language models (LLMs):


Another really good comparison was provided in the article, What is the relation between Semantic Web and AI?, which contrasted rules-based artificial intelligence and statistical machine learning:


The hybrid approach on the RIGHT combines the features of both rules-based systems and statistical machine learning. This approach is referred to as neuro-symbolic AI it appears.

To be able to properly differentiate the true capabilities of each approach, it is very important to understand the difference between deductive reasoning, inductive reasoning, and abductive reasoning.
This short video, Deductive vs Inductive vs Abductive Reasoning, helps you distinguish between the three approaches.
  • Deductive reasoning: Start with a general rule, apply the rule to a specific situation; the result will always be true (i.e. there is certainty). In deductive reasoning you start with one or more preemies which then lead you to a conclusion.  As long as the fact are right and you follow the rules; then the result will always be certain. 
  • Inductive reasoning: Start with a specific observation, reach a general conclusion based on that available evidence, reach a general conclusion that may or may not be true. (i.e. the results can never be for certain).  Inductive reasoning is about making repeated observations and then reaching a conclusion based on those observations.  There is always a possibility that the conclusion can be wrong. Inductive reasoning can be very useful even though results are not certain.
  • Abductive reasoning: Start with an incomplete set of observations, use that information to make a plausible prediction as to the right answer which may or may not be true (i.e. the results can never be for certain). Abductive reasoning is about making a well-informed assumption based on the evidence available; it is a prediction. Abductive reasoning can likewise be very useful even though results are not certain.
Logic is the study of correct reasoning. Logic is a formal system of reasoning. In a logical system something cannot be both true and false at the same time. Informal logic (a.k.a. logical fallacies or fallacies of critical thinking) is about things that may seem logical to someone not trained in formal logic; but they are not actually logical.  Critical thinking is the analysis of available facts, evidence, observations, and arguments in order to form a judgement by the application of rational, skeptical, and unbiased analyses and evaluation.  Logical thinking (a.k.a. formal logic) is the process of evaluating truth conditions and the legitimacy of connections between logical statements by applying formal deductive logic. Critical thinking tries to employ the ideas of logic but it is less ridged that formal logic.

To understand the true capabilities of artificial intelligence, to set the appropriate expectations of these tools, and to interpret the results these tools appropriately; one needs to have a precise understand as to how these tools actually work.  With machine learning there can always be Black Swan events.  When used appropriately, these can be very powerful and useful tools.

Artificial intelligence will undoubtedly supercharge the capability of many software tools.  But be sure to use the right tool for the job.

Additional Information:

Comments

Post a Comment

Popular posts from this blog

Relational Knowledge Graph System (RKGS)

Image
Relational Knowledge Graph System (RKGS) is an interesting idea that combines the relational paradigm and the knowledge graph paradigm.  This documentation, Why RKGS , explains the idea. This gets even more interesting with the notion of "relational AI".  RelationalAI is described as follows by the company of the same name, Relational.ai : “RelationalAI is a cloud-based relational knowledge graph management system, with state of the art probabilistic processing and declarative reasoning at scale to make developing Data Applications a superpower for your business.” This notion of knowledge graphs represented in relational databases fits into my belief that knowledge graphs have two aspects: the expression SYNTAX and the logic of what is expressed or SEMANTICS. While the LOGIC of a knowledge graph is always the same regardless of technical syntax used to express that logic; there are many, many different technical syntaxes that can be used to represent a knowledge graph. There
Read more

Getting Started with Auditchain Luca

Image
(Note that a newer, production version as opposed to only the development version , of Luca Suite is available) The Innovator's Dilemma points out that disruptive innovations often look like toys that are doing everything wrong. Auditchain Luca , which is part of the Auditchain Suite , is a unique approach to creating a knowledge graph .  To experience using Auditchain Luca, have a look at the Getting Started with Auditchain Luca instructions. Be sure to grab the Excel and JSON import files used for this getting started exercise.   This getting started exercise uses one of several different approaches to getting information into a tool for constructing and viewing professional quality knowledge graphs .  Additional getting started guides will be made available over time. Auditchain Luca might not make sense to you if you don't understand the paradigm shift that is occurring and if you have not shifted your mental map. For the next step in your journey, consider trying creat
Read more

Evaluating the Quality of XBRL-based Financial Reports

Image
From the time public companies started submitting XBRL-based financial reports to the U.S. Securities and Exchange Commission, I have been looking at those reports, trying to understand the XBRL-based reports, and trying to figure out how to get them right.  That information was used to figure out how to create software applications that help professional accountants get these reports right. Here is information that summarizes my testing and results which I stopped doing March 31, 2019: Analysis of the high level accounting relationships in the 10-Ks of about 5,716 financial reports  (5,063 public companies, about 89.1% got all these high level accounting relationships consistent with expectation, 623 were inconsistent and manual confirmation revealed the inconsistency was an error) Documentation that helps you understand 26 different types of errors in XBRL-based reports High quality documentation of hundreds of errors in XBRL-based reports submitted by public companies to the SEC Doc
Read more