What is AI?
Some informal definitions and categories
In the last blog post, I talked about the need for knowledge representations to capture human experience and concerns, and the need for AI research to include this. AI has been been talked about a lot lately, but mostly from the point of view of machine learning and generative AI (such as ChatGPT). Other areas of AI have received much less coverage (such as symbolic reasoning). I will now discuss what is meant by AI in this broader sense, and show where knowledge representation fits in. My goal is to use non-technical language (as far as possible) while also using precise terminology, in the sense of being unambiguous. For deeper coverage, I would recommend the book Artificial Intelligence: A Guide for Thinking Humans by Melanie Mitchell.
A key take-away of this post is the importance of human experience and human authorship in the building of AI systems.
What is Intelligence?
AI is sometimes defined as designing machines to do tasks that would require intelligence for humans to do. But “intelligence” is difficult to define. The concept of “IQ” as a measure of “general intelligence” has always been controversial, although it can sometimes be useful. (See for example, this article by Daphne Martschenko). Artificial General Intelligence (AGI) is also a disputed concept, and I find it vague. Similarly, “human-like” or “human-level” intelligence is hard to define, so I try to avoid using these terms, as well as “AGI”.
Intelligence can be viewed as having multiple dimensions, some of which can have quantitative measures. So for example, language and mathematical abilities could be measured numerically on two different dimensions (because it is possible to be good at one, but not the other). But there are other kinds of intelligence that are more difficult to quantify, such as social and emotional intelligence, and also creativity. So I prefer to think of "human intelligence” as a collection of capabilities and not a single number.
AI as Automation
I like to think of AI in terms of automation. Traditional automation is about automating physical processes. AI is about automating mental processes, such as reasoning, decision-making, and pattern recognition. (I am excluding emotion and consciousness for now. They require a separate discussion).
Instead of talking about “AGI” or “human-level” intelligence, I think it is helpful to define capabilities that an artificially intelligent system needs in order to automate tasks reliably. So for example, logic-based reasoning is a useful capability for decision-making and planning. The capability to handle noisy and changeable environments is important in artificial vision.
Categories of AI
AI is a very broad research area, with many different approaches. Two of the most important are as follows:
Statistical AI: emphasises statistical approaches to pattern-recognition and machine learning (ML). This includes neural networks which are based very roughly on brain models.
Symbolic AI: emphases human knowledge and meanings. This is sometimes called “Good Old-Fashioned AI” (or GOFAI) - but I think it is still necessary.
Statistical AI
This approach is focussed on pattern recognition and adaptation in environments which contain unstructured and noisy data. It includes statistical machine learning (ML) and generative AI. The aim is to produce responses or decisions that are good approximations of the required answer. Examples of problems solved by statistical ML include identifying spam email and detecting financial fraud by analysing transactions.
Generative AI includes Large Language Models (LLMs), such as ChatGPT. An LLM produces text outputs that are similar to typical answers to the question in a prompt. Because of the massive amount of training data available, it can learn the pattern of answers to common questions.
Neural networks are the key technology behind machine learning and generative AI. They are networks of nodes (representing “neurons”) which are gradually adapted to recognise patterns in unstructured data, such as text or images. (Although they are called neural networks, they are extremely simplified in comparison with real brains). Neural network-based AI has become very successful lately because of the advancement in computational processing power and the massive amounts of data available.
Symbolic AI
Human mental processes often involve the transformation of symbolic content using reasoning. For example, if we are informed that all rooms in a particular hotel have air conditioning, and a room is available in this hotel, then we can reason that the available room has air conditioning. Similarly, we understand negation statements such as “smoking is not permitted in any of the rooms”.
Symbolic AI is the automation of human symbolic content processing. The simple kinds of reasoning above can be automated, provided the relevant true statements are defined in advance, along with rules for generating new statements. This kind of automation doesn’t include the nuance and imagination involved in meanings, but it can provide reliable reasoning that is good enough for many kinds of problem-solving. Importantly, it can give a trace of the reasoning steps used to arrive at a decision, which enables transparency.
Symbolic vs Statistical AI
Symbolic AI alone has serious limits due to the need for manual specification of all the facts and rules. For this reason, applying it to real-world problems has been difficult (because the world is messy and unpredictable). Interestingly, there are purely symbolic approaches to machine learning. (See for example, this informal overview of Explanation Based Learning, which produces general rules by reasoning about specific examples). But these symbolic methods are limited because they need detailed prior knowledge. However, the currently popular approaches to machine learning and generative AI also have disadvantages. Some of these are discussed below.
Reliability
Symbolic AI systems can produce reliably correct answers to logical reasoning problems where the facts and changing states of the world can be precisely defined. So they are good at games and puzzles such as the river crossing puzzle. By contrast, LLMs are not guaranteed to give the correct answer to logic problems because they do not apply the rules of logic, but instead generate a typical response to the question in the prompt. If the question is a common one, the answer is probably correct in most cases. But if it is rare, the answer is often wrong. See for example, this discussion on the limits of reasoning in LLMs.
Transparency
Although the outputs of a statistical AI system may be in a symbolic form (such as text), the steps by which it arrives at the output are very different from human thinking. For example, when an LLM chatbot answers questions, the inputs and outputs are symbolic. The input is a question in natural language (such as English) and the output is in the same language using symbols that are meaningful to humans.
But statistical AI systems do not use internal symbols that can be interpreted by humans. Instead the words in the input prompt are converted into rows of numbers, which are then processed numerically. The output text is generated by predicting the next word. Therefore, if such a system makes a decision or generates a text output or image in response to a prompt, it is difficult to understand the processing that led to its response. This is the AI transparency problem. There are no internal symbols or transformations of symbolic content as there is with symbolic AI. For this reason, “explainable AI” (XAI) has become an active research area.
Knowledge Representation and Human Authorship
Returning to symbolic AI, facts and rules are not the only way to represent knowledge, but they are a simple place to start. They can be used to capture concepts and concerns, which are key components of mental models, as discussed in the last post. Representing knowledge means that choices must be made about what facts/rules are important. The key point here is human authorship. The choices about what to include and how to express it are based on human experience - which can be everyday lived experience or subject-matter expertise. In other words, machine learning cannot do everything. I plan to discuss knowledge representation in more detail in the next post.
Notes
This blog post is human-produced, without any generative AI tool.