This is part of a four post series spanning two blogs.
- One post gives a general historical overview of the artificial intelligence business.
- One post specifically covers the history of expert systems.
- One post gives a general present-day overview of the artificial intelligence business.
- One post (this one) explores the close connection between machine learning and (the rest of) AI.
1. I think the technical essence of AI is usually:
- Inputs come in.
- Decisions or actions come out.
- More precisely — inputs come in, something intermediate is calculated, and the intermediate result is mapped to a decision or action.
- The intermediate results are commonly either numerical (a scalar or perhaps a vector of scalars) or a classification/partition into finitely many possible intermediate outputs.
Of course, a lot of non-AI software can be described the same way.
To check my claim, please consider:
- It fits rules engines/expert systems so simply it’s barely worth saying.
- It fits any kind of natural language processing; the intermediate results might be words or phrases or concepts or whatever.
- It fits machine vision beautifully.
To see why it’s true from a bottom-up standpoint, please consider the next two points.
2. It is my opinion that most things called “intelligence” — natural and artificial alike — have a great deal to do with pattern recognition and response. Examples of what I mean include:
- Think of what’s on an IQ test, or a commonly accepted substitute for same. (The SAT sometimes substitutes.) A lot of that is pattern recognition.
- When the “multiple intelligences” or just “emotional intelligence” concepts gained currency, the core idea was the recognition of various different kinds of pattern. (E.g., reading somebody else’s emotions, something that I’m not nearly as good at as I am at the skills measured by standard IQ tests.)
- The central mechanism of neurotransmission is a neuron recognizing that an action potential has crossed a certain threshold, and firing as a result.
- Traditional areas of AI include natural language recognition, machine vision, and so on.
- Another traditional area of AI is rules-based processing — conditions in, decision out.
- Back in the 1980s (less so today), it was thought that a core underpinning for AI technology was knowledge representation. That said, as much as I like interesting data structures, I have my doubts.
- The Semantic Web grew out of this idea.
- Also, the single most enduring proponent of the centrality of knowledge representation was probably Doug Lenat, who gave his name to a famed unit of bogosity.
- While the previous two points are probably just coincidence, the juxtaposition is suggestive.
![:)]()
3. In most computational cases, pattern recognition and response boil down to scoring and/or classification (whether in a narrow machine learning sense of “classification” or otherwise). What I mean by this is:
- I’m thinking of scoring as a function that maps inputs into scalar values. (Or a vector of scalars.)
- I’m thinking of classification as a function that maps inputs into a finite range of possible values. (Note that this is mathematically equivalent to a finite partition on the set of inputs.)
- I’m also assuming that the system maps each possible score or classification to a decision or response (deterministically or probabilistically as the case may be).
- Then if you compose the two maps, you wind up with a function from {possible input patterns} to {possible responses}.
4. If you want a good algorithm for classification, of course, it’s natural to pursue it via machine learning. And the same is true of scoring, at least if we recall that the domains of machine learning and statistics have essentially merged.
5. It took people remarkably long to figure out the previous point. Through at least the end of the previous century, it was generally assumed that the way to come up with clever algorithms for, for example, text analytics or machine vision was — well, to think them up.
6. As spelled out in my overview of present-day commercial AI, there’s a somewhat paradoxical industry structure, in that:
- Even though machine learning is a sine qua non of many businesses, tech and non-tech alike …
- … the rest of AI is largely concentrated at a few behemoth technology companies.
Of course, there are plenty of startups hoping to change that structure. I hope some of them succeed.