Ice cream and the folly of machine intelligence “truly understanding”
There’s a big debate at the moment in the natural language processing community about whether deep learning can create “true understanding”.
Gary Marcus tweeted recently:
This got me thinking about what this means for “true understanding”, and how I’ve seen my kids gain understanding of the world. Kids copy what you say. And first off, they just parrot what you say, with no understanding of what they mean. Over time, when they say things, what they say can be more coherent because the words they use are linked to other knowledge (words, experiences, images, and other sensory experiences).
Brilliant example from a year ago:
My 7yo: “Daddy can I have some I-C-E C-R-E-A-M?”
Spelling words out is a kind of 7yo superpower that allows him to communicate with other people in a kind of “code” about things that he didn’t want his little 3yo brother to know about.
My 3yo, however, noticed something was up:
“Daddy can I have some I-C-E C-R-E-A-M?”
For non-parents, no, my 3yo is not a prodigy who learnt to spell at 3.
So what happened? They literally said the same thing. What almost certainly happened is that my 3yo observed his big brother and realised that he was asking for something, and that something would probably be something he too would want. So he asked for exactly the same thing, parroting the exact syllables but not having any idea that:
- they were representing letters,
- that they were spelling a word, and
- that that word in fact was something he did in fact like! Ice cream!
I’m sure there is a huge body of knowledge around child development that has all this covered well, and I look forward to learning more formally about this.
When thinking about machine intelligence, I couldn’t help marvelling at this amazing ability children have to learn by example, progressively refining, correcting, and, critically, deepening understanding of the world and language to represent it and its relationships. Kids just imitate, mimic, and copy, they see how the world responds, and they learn about how that action impacts the world. In this case, if an ice cream turned up, there’s a fairly good chance that my 3yo would start linking the act of asking about the spelling of ice cream with receiving some. However that link needed to know nothing about spelling or words! That might come later if it was useful.
And so at what point do we humans really say we “truly understand” what’s being said? My examples above suggest that understanding is a continuum. This of course seems obvious when looking at human behaviour, but coming from a technology perspective, people tend to be a little more black and white. “ah yes but did the computer truly understand what is being said?”.
As such I believe that aiming for AI that “truly understands” is folly. And the most useful is in fact, for a given context and use case, whether it understands to the depth necessary to serve its purpose.
