Artificial Intelligence: You know it isn't real, yeah? "Where's the intelligence?" cried a voice from the back. It's not quite the question one expected during the Q&A session at the end of the 2019 BCS Turing Talk on Artificial Intelligence. The event was held earlier this week at the swanky IET building in London’s Savoy Place and the audience comprised academics, developers and …
Hi to all, and, the error comes not just from "Joe Average" but from most of "Tech" people too ...
It is NOT AI, is is Machine Learning !!!!
We do NOT have AI, nor will have in any foreseeable future ...
What we have now, and for quite some time now, is Machine Learning .
Until people use the wrong terminology, the understatement of it will be wrong ...
Even ML isn't really right. The machines aren't learning, we're aren't "teaching" them to do anything. All we are doing is using datasets to produce statistical predictions on future data. Model wrong? Bad data? Start again from scratch, because the prediction outcomes will be wrong - the machine doesn't learn anything from the previous incarnation.
Automated stereotyping is actually an excellent description. We should switch to it immediately.
The goal is to classify (or otherwise map) a large and variable data set as lazily and efficiently as possible. So the NNs do the same thing we do -- pick some easy to spot things (proxies) that are correlated in the cases encountered so far. Because when we state the problem like that, this is simply the solution (in our case we arrived to it by evolution). Except, unlike us, the poor NNs cannot reason about it. Not that we do it often, but anyway...
"
So the NNs do the same thing we do
"
Except we have not mapped only one or two data sets, but many hundreds of data sets. Often we need to make use of correlations learned from one data set and apply it to a situation that is normally the domain of a different data set.
For example, how we react while driving and an unexpected object appears in the road depends whether the object we see maps into the data set of "very hard solid things" (e.g. big rocks), "Soft inanimate things" (e.g. a bin bag blowing in the wind), "Small animals" (e.g. fox), "possible human" (e.g. runaway pushchair) or "Something abnormal" (e.g. very big pothole, open fissure, collapsed drain etc.). We also react based on our knowledge of human behaviour - e.g. a ball bouncing across the road is itself no danger, but may well be followed by a child chasing after it.
We can also recognise and appreciate the difference between a road that has a soft verge, a road that has a ditch next to it, and a road that has a sheer 200 foot drop next to it, and this will influence our decision on the best course of action to take in order to try to avoid a collision.
The "A.I." in a driverless car would be able to recognise a tiny fraction of the things we are able to recognise, because its "knowledge" might be quite deep, but is nowhere near as broad.
Plus, some of the things we do are done without our conscious knowledge of it. I've you've ever played QWOP or derivative games like Manual Samuel), you start to realize how much motion we make without actually giving thought to how we move. It's been said the hardest thing to program will be intuition, simply because we don't understand it ourselves as it's all subconscious. And intuition may just be the mechanism behind "inspirations" that take seemingly unrelated things and apply them in completely novel ways to form a solution that we may not have even been consciously seeking. Forget artificial intelligence. We still don't have a full grasp of intelligence, full stop.
We can also recognise and appreciate the difference between a road that has a soft verge, a road that has a ditch next to it, and a road that has a sheer 200 foot drop next to it, and this will influence our decision on the best course of action to take in order to try to avoid a collision.
Some humans can, but I find that living in a mountainous region with twisty roads, vertiginous drops and deep ditches on the uphill sides results in loads of ruined tyres when you meet f**king tourists who freeze in the middle of the road and force you to reverse 200m round 2 blind corners until the oncoming driver thinks the road is wide enough to pass, instead of just moving to their own side of the road and letting you drive past. I'll admit some of the bridges are narrow enough that at least one car has to retract the wing mirrors, but that is no excuse for staying in the middle of the road to prevent normal passing. Even a car driven by an ML system could work out where the edge of the road is.
"Some humans can, but I find that living in a mountainous region with twisty roads, vertiginous drops and deep ditches on the uphill sides "
Sadly, that is not restricted to where you are. Many, many drivers have no idea how wide their car is or have any concept of where the nearside of their car is despite the many situations where the road narrows in town, eg cars parked on both sides of the road.
"The "A.I." in a driverless car would be able to recognise a tiny fraction of the things we are able to recognise" - Sorry but selfdriving vehicles are able to recognise much more than human and have a lot more possibilities to crunch a lot of "impressions" in a very short time. So every selfdriving car knows that if a ball crosses the road, there might be a child that follows. Maybe it knows it very well because the car in the front has recorded this child and already informed the following cars. The car also recognizes if someone standing beside the road is shaking his head and maybe knows that the person did run for the last few minutes because he/she wants to take the bus that will stop on the other side in just one minute. A selfdriving car knows what is happening after a hill what a human never could know/see. So it is not about one car. It's about the connection and reaction to a lot of data, delivered by a lot of devices (cameras, web, smartphones, smartwatches, fitnessbands, traffic lights etc.) much more data that a human can analyse in the same amount of time.
Perhaps one day but we simply do not have systems that 'know' a ball may be followed by a child. That particular scenario may well be currently 'coded' for but even recognising a ball is a tough task in many conditions (for machines that is).
We have not crossed (and may never cross) the threshold where the machine 'thinks' such that it can determine actions well in new situations dissimilar to previous ones.
It is true that a future system (including the car, other cars, roads, traffic lights, 'bend sensors' (tm)) may well yet make us incapable of killing ourselves or others with vehicles.
But not yet and not for some time.
"
Sorry but selfdriving vehicles are able to recognise much more than human
"
Is that why one failed to spot a huge frigging articulated lorry parked across the road?
Sorry but an "AI" can only recognise things and situations that it has been programmed to recognise. Humans (by the time they can drive) have had at least a decade being programmed to recognise a a far greater number of objects, object-sets and complex situations.
The machines aren't learning, we're aren't "teaching" them to do anything. All we are doing is using datasets to produce statistical predictions on future data
Would you care to articulate and support your theory of learning which demonstrates a substantive difference from a statistical process that makes "predictions on future data"? Feel free to draw on logical positivism, metaphysics, phenomenology, pedagogical theory, cognitive science - I'd just like to see an actual fucking argument rather than handwaving and posturing.
(I'll ignore for now the fact that "produce statistical predictions on future data" is not an adequate or useful description of ML.)
"Whatever happened to the term "expert system"?"
From what I remember an expert system made decisions using rules defined by human subject matter experts. Usually those rules were associated with reasons, so an expert system could not only make a decision but also support that with the underlying reasons.
AI/ML (which I know little about) uses statistical analysis to discover correlations. However, most of us will be aware that correlation <> causation. Hence it might be safest to use AI/ML as a tool to discover interesting associations which can then be considered by humans. Furthermore I get the impression that AI/ML cannot give reasons for the decisions it makes/recommends. In my view that makes it unacceptable as a decision-making tool (though it may be a useful decision-support tool).
"Usually those rules were associated with reasons, so an expert system could not only make a decision but also support that with the underlying reasons."
The shortcoming about that is that is can only make the decisions it was given reasoned rules to make.
A real expert, OTOH, will have a degree of understanding that helps them, when confronted with a novel situation, to undertake new reasoning and at least suggest what the best decision might be.
AI works with connected by their sense paragraphs, where each paragraph contains a (more or less) clear formulated and described idea. Described in terms of all weighted the paragraphs' patterns, where these weights are the only AI statistics.
These paragraphs are derived from (more or less) meaningful texts that (sometimes, often) contain causal and logical connections. AI captures this relationship by taking into account their timestamps and establishing relationships between patterns of paragraphs.
Initially, AI technology was developed as a response to the challenge of NIST TREC QA - how to find answers to Factoid and Definition questions? The problem of explaining " What, How and Why?" was not important - I need to find answers.
Would you be so kind as to take Merriam Webster? Look, after the definitions often are given examples of their use. Do you see?
Structuring paragraphs you create and add (to the definitions from the dictionary, for the words of the paragraph) examples of their use. That is, you create uniqueness, tuples for the words of the paragraph, and the more accurately expressed the idea in the paragraph the more accurate its tuples.
And then the word and its phrase can be easily found. For example in NIST TREC QA there were more than 6 million texts in which it was necessary to find one word/ phrase in its sense. It is obvious that without the structuring of paragraphs it's impossible!
According to NIST TREC the only system that can answer (Factoid and Definition questions) is considered to be the true AI, which gave me the right to claim that I created AI.
It was all the rage in the 80s but fizzled away. Now marketeers need a new label to sell stuff based on that.
On a side note: I was in a seminar about "Data Scientist being the Sexiest Job in the 21st century". We jokingly concluded that the next fad sexiest new promising job title will soon be "Data Engineer" (6,430,000 hits in Google) then "Data Architect" (3,680,000 hits in Google) then "Data Consultant" (596,000 hits) "Data Decorator" (10,500 unrelated hits) then "Data Feng-Shui Consultant" (no hits??) then probably "Data Psychic".
GIGO - whoever has control of the datasets has control of the output, and there isn't a single data set that isn't biased by some rule(s) that has to be decided upon arbitrarily. Someone will round to whole numbers, for example, while someone else might round to 10ths, and someone else to 100ths - something that simple when run through any algorithm will create more and more distortion over time. Then there is the guy that, in setting up a simple data set (what data needs to be imputed to make sure an airplane will make it from point a to point b) assumes that the 20000 represents gallons, rather than pounds, of fuel), and the plane lands 2000 miles short of the runway.........
Pretty sure that when i were a lad, this AI stuff was being sold under the "cybernetics" buzzword, and personally that makes more sense to me... the greek steersman seems a much more realistic image than the godlike world-mind that popular fiction associates with AI
Trouble is, cybernetics now points to concepts like cyborgs (in the strictest sense, living beings augmented with nonliving components, which is something that is actually in progress) so the word now invokes images of a more direct man-machine relationship.
this AI stuff was being sold under the "cybernetics" buzzword
It may well have been, but that's a meaningless characterization. "Cybernetics" means one of two things:
1. The technical term Norbert Wiener invented for the scientific study of self-regulating systems, from the Greek kybernetes, meaning "steersman", or various cognates related to governance.
2. A largely nonsense term in popular discourse with no well-defined meaning, other than vague connotations of mechanization and information technology.
ML and the broader category of AI have applications in cybernetics (as do many unrelated technologies, such as fuzzy logic and systems theory), but in no way are they equivalent to cybernetics.
From experience of living in the area it would seem that a modicum of Chelsea supporters clearly demonstrate every week or so that whatever ‘intelligence’ they possess can only be determined as artificial. Someone clearly removed any intelligence that they once possessed and replaced it with a limited capability to converse in simple words such as ‘beer’, ‘yer what’ and the occasional strung together sentence such as, “I hate Tottenham/any team from Manchester”.
In fact the typical Chelsea supporter is an example of a system that can run two programs in different environments but can't multi-task. During the week they are solicitors, bankers and chartered accountants; the proximity of a football ground causes an environment switch to yob mode.
Group mentality can be blamed on evolution, though I fear it could be a behaviour I should be thankful for. Allowing our early ancestors to consider other groups /species as the enemy and to kill or steal from them, just because "Oh fuck, they're getting better than us".
"Data analysis and pattern matching" would be a better term, as it doesn't learn anything.
Again: This is a vapid claim if it's not accompanied by an at least moderately rigorous definition of "learning".
ML systems accumulate information entropy, and use it to constrain the range of their output. That's a pretty supportable definition of "learning".
Each pattern is a direct analogue of a programming language command, in its unique subtext and context. AI learns by addition of new patterns, subtexts and contexts.
For example, there is a paragraph which has a pattern (phrase, a few words). The paragraph and surrounding it paragraphs dictate the pattern context and subtexts; where each dictionary definition - for each the phrase's word - is its one subtext.
I totally agree but it is seemly impossible to get academics, business, governments, media and the world from using "AI". Today's so-called AI is not what McCarthy was describing with term in 1956. It is, as you say, Machine Learning and we should expect this "learning" to be very lumpy. We should also take a look at the rise of statistics and probability theory in the 19th and early 20th centuries. Many of the difficulties ML is generating, including the hype and myth, were first generated over a century ago. ML has automated these analytical methods which of course adds another turn of the screw.
Hopefully the second wave of AI hype is beginning to subside and we can sensibly investigate the potential and limitations of the technology. I do have my doubts about this as there is too much invested in the myth of AI.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
