The Platonic Representation Hypothesis

Phillip Isola, MIT

[doublepinger]

The part at the end about a blind person being given sight - as far as I'm aware, blind people given sight still end up using their cane. The "simple" obstacle of a curb is very foreign to them, by sight.

[KiriIrving]

The discussion around 48:30 is literally textbook perfection example of the scientific method.

It's an observation, question, and testable hypothesis that follows from your understanding of the topic, but has not yet been tested. *chef's kiss*

[angelorf]

Very clear lecture. Thanks for this!

Seems to me you are throwing a lot of engineering and science against a problem which is fundamentally philosophical. Assuming that our senses observe aspects of the real world is a necessary prerequisite for doing science and it is similarly a prerequisite for machine learning to have any use. It doesn't require empirical validation.

The hypothesis aligns better with Aristotles philosophy than with Platos.

If you ask me to which image the snake source image is more similar, the black snake or the closeup of a snake then I would say your question is unclear. What does that experiment tell us about how humans tend to interpret an unclear objective?

Very nice work at the intersection between my two studies: philosophy and artificial intelligence.

[mslapik]

As humans, we've built physical monuments to represent our abstract ideas. Concepts like universities, companies, and countries—while abstract—are embodied by physical structures in the world. Could a vision model, then, use these physical objects to learn about the abstract concepts they represent?

[rysw]

This is extremely interesting research, but the conclusions are fundamentally flawed. We build these systems with all kinds of priors based on our own capabilities, then we judge their effectiveness in relation to our own objectives.

The models are reflections of ourselves first and foremost, and any relation they have to the world is mediated by the connection we have with the world.

[richardnunziata]

Has attention and saliency been consider in these

[blengi]

hmmm, this AI stuff seems a bit isomorphic to my information agnostic abiogenesis matrix model, which for various reasons kind of implies exist an ecology of universal latent space representations via auto selection over entropic information compactification.. Even hypothesized similarly in an AI LLM context about _platonic information representations forms convergently arrived at_ elsewhere ie before this paper. Maybe they read lol?

[MalachiSpring-st]

Gonzalez Sharon Hernandez Jessica Hernandez Edward

[HopkinsDean-ri]

Walker David Rodriguez Timothy Wilson Robert