Me third go at a zero shot, by analogy metaphor carried me here.
Me third go at a zero shot, by analogy metaphor carried me here.
wherein the stochastic parrot is not mentioned
When a “trained model”, i.e. an AI model which has ‘learned’ to classify some lumps in our world, is then asked to classify new sets with lumps of our reality that it has not been trained on directly, but has been exposed to indirectly in another algorithmed dataset (i.e. it can recognized traffic lights, but could it class shadow puppets based on their descriptions in Javanese traditional textual works?) has been called zero-shot.
According to Wikipedia this was originally called dataless classification, but the play on one shot in one-shot learning that was introduced in computer vision was too cool to avoid, and it already had currency… — also dataless classification is way too generic to stick IMHO.
Basically they were mucking around and mashing things up and found something interesting.
(A) When I first heard this term zero-shot I did not get any of this context. I just ‘heard’ that it ‘guesses’, somehow, but I was assuming a general intelligence that happens upon something novel. And not an indirect cross-infection (which I suspect is the result of worlding commonalities, animal, human or mineral.)
I was using the wrong analogy (assuming a general intelligence ).
If only because a general intelligence does not exist yet, at least artificial general intelligence (AGI) does not exist as I had somehow slipped into though.
The question of analogy is key here.
I am writing about how I learned about the jargon used in artificial intelligence research and development.
I am not writing about general intelligence.
Except I am.
In a way.
I guess.
I.
What is the mechanism of analogy?
When a model engages ‘successfully’ in zero-shot classifications, it is uses patterns in an intermediary dataset that humans have originally created as animal intelligences, in meatspace. It reflects our world, our shared umwelt. It means patterns picked out by the intermediate model may apply in a third model where the zero-shot takes place.I’ve counted up to at least three here, thus my confusion mentioned up at (A) that this three is not even one. It’s our experience and intent that is extended, it is the measure of us.
It is important to note the the intermediate dataset was not necessarily used to train the AI model on Y, but contains Z as constrains and contexts within Y… —somehow. Thus its ability to zeroshot comes out of co-founded classes unintentionally collected by the training.
It only seems zero because like you ain’t used it there before.
This is not general intelligence. This is an artefact of the world we live in.
I guess we humans may not have general intelligence either.
I’ll write about analogy/metaphor and general intelligence another day.
