[HN Gopher] Clever Hans (Intelligence Misattributon)
___________________________________________________________________
Clever Hans (Intelligence Misattributon)
Author : anotherhue
Score : 52 points
Date : 2023-03-26 17:28 UTC (5 hours ago)
(HTM) web link (en.wikipedia.org)
(TXT) w3m dump (en.wikipedia.org)
| StrangeATractor wrote:
| Sounds like Elon Musk fans.
| LesZedCB wrote:
| emotional intelligence is intelligence, too. sheesh. the premise
| is flawed because the horse was engaged in a complex intelligent
| interaction with the person.
|
| trickery aside, Hans was clearly intelligent by several metrics.
| mathematics are not the sole domain.
| PaulDavisThe1st wrote:
| where do you see any suggestion that Hans lacked all
| intelligence?
| LesZedCB wrote:
| "Intelligence Misattribution"
|
| right in the title.
|
| either it's clearly a gradient/spectrum, or this article
| isn't making the solid argument it might seem to be.
| PaulDavisThe1st wrote:
| no, it's not a gradient/spectrum. The HN title is just
| missing extra words, like "Mathematical".
|
| Nobody said that Hans had no intelligence, the point is
| that Hans doesn't have the kind of intelligence that was
| being attributed to him from the arithmetic demonstrations.
| theknocker wrote:
| [dead]
| version_five wrote:
| I actually just posted an article that references Clever Hans
| (pretty sure it's also the name of a Grimm Brothers fairy tale,
| most of them have names like that).
|
| It's actually not the best analogy for the current fad, I think
| "confused by a mirror" works much better.
| https://news.ycombinator.com/item?id=34850890 although I suppose
| the idea of attributing intelligence where there isn't any holds.
|
| Clever Hans is a good analogy for (mostly deep learning) ML
| classifiers, the idea being a, say image classifier, can do
| extremely well despite not using any features that actually
| causally inform the decision being made. So when the background
| changes, or lighting, or some other thing, it fails utterly. This
| is actually true for almost all image classifiers.
| anotherhue wrote:
| I generally agree, Hans came to mind when I read a recent post
| where people discovered they had been nudging gpt towards
| answers without realising it.
|
| https://news.ycombinator.com/item?id=35306095
| version_five wrote:
| Ah, interesting. I suppose that's a lot like what
| psychologists or psychics try and do, convince you to prompt
| them without you noticing.
|
| Reminds me a bit of an old demo where you're trying to
| measure a certain signal shape in the presence of heavy
| noise. You get shown waveforms and decide if it looks like
| there might be a signal present or not, and the ones you
| approve get averaged to pull the signal out. Very quickly,
| even if all you are shown is zero mean gaussian noise, the
| average of what you OK'd starts to look like the waveform
| you're looking for.
| Tao3300 wrote:
| Eisenhans (Iron John)
| kune wrote:
| "Der kluge Hans" is not directly a Brother Grimm title. There
| is a story "Der kluge Knecht", which translates into "The
| clever farmhand". The hero of the story is called in the story
| kluger Hans. There is another story called "Der gescheite
| Hans", which means exactly the same. Both stories are not fairy
| tales but droll stories about a simple-minded person.
| ksherlock wrote:
| The English translation includes "Clever Hans" which seems to
| be Der gescheite Hans.
|
| https://www.gutenberg.org/cache/epub/2591/pg2591-images.html.
| ..
| seydor wrote:
| Intelligence misattribution does not follow from the story of the
| horse
|
| More like accidental trickery
| glutamate wrote:
| GPT-4 is Clever Hansing at web scale
| misnome wrote:
| I think of it more like extremely efficient large-scale Cargo
| Culting. It knows exactly what it should look like, without any
| conception of why or the originating factors that led to it.
| jszymborski wrote:
| I only knew of this story from looking up the name of this
| library on adversarial DL https://github.com/cleverhans-
| lab/cleverhans
| lostmsu wrote:
| Title is editorialized.
|
| It is also inapplicable to GPTs, as they can answer questions,
| which the person asking does not know the answer to.
|
| Case closed.
___________________________________________________________________
(page generated 2023-03-26 23:01 UTC)