platonic representation hypothesis
In plain wordsThe idea that bigger, more capable AI models, even ones trained on different things such as images and text, end up organizing information in increasingly similar ways, as if all were mapping the same underlying world.
The hypothesis that, as neural networks become more capable, their learned representations grow more alike regardless of architecture, training objective or even data modality, approaching a common statistical description of the world that produced the data. Evidence comes from representational-similarity measures, chiefly a mutual-nearest-neighbor kernel-alignment metric: larger, better-performing vision models agree more with one another, and language models' representations of captions align increasingly well with vision models' representations of the matching images as both improve. The proposed explanation is that many tasks, large capacity and a bias toward simple solutions jointly push representations toward the same solution. Critics question whether the alignment metrics capture meaningful agreement, note that convergence may reflect shared data and objectives, and point to persistent differences between models. The idea extends the universality hypothesisuniversality hypothesis, n.: The idea that different AI models trained on similar tasks tend to develop the same internal concepts and mechanisms, so what is learned about one model may carry over to others. from individual features to whole representations.
If the platonic representation hypothesis is right, a text model and an image model should agree on which concepts are neighbors.
Origin
Proposed by Minyoung Huh, Brian Cheung, Tongzhou Wang and Phillip Isola of MIT in 'The Platonic Representation Hypothesis' (May 2024; ICML 2024). The name alludes to Plato's allegory of the cave, with training data as shadows of an underlying reality.
Related
- universality hypothesisuniversality hypothesis, n.: The idea that different AI models trained on similar tasks tend to develop the same internal concepts and mechanisms, so what is learned about one model may carry over to others.
- representation learningrepresentation learning, n.: Letting a computer work out for itself how best to describe its data, such as the edges and shapes in images or the meanings of words, instead of relying on descriptions people design by hand.
- embeddingembedding, n.: A list of numbers that stands for a word, picture or other item, arranged so that similar items get similar lists. It lets a computer measure similarity of meaning as distance.
- CLIPCLIP, prop. n.: An OpenAI model that learned to match pictures with their captions from hundreds of millions of web examples, so it can judge how well any piece of text describes any image.
- linear representation hypothesislinear representation hypothesis, n.: The idea that a neural network stores many concepts as simple directions inside its internal activity, so how much of a concept is present can be read off by measuring along its direction.
Sources
- Huh, M., Cheung, B., Wang, T. & Isola, P. (2024). The Platonic Representation Hypothesis. ICML 2024. arXiv:2405.07987. arxiv.org/abs/2405.07987
Mentioned in
centered kernel alignmentcentered kernel alignment, n.: A score from 0 to 1 measuring how similar two neural networks' internal representations are, by comparing how each arranges the same set of inputs relative to one another., universality hypothesisuniversality hypothesis, n.: The idea that different AI models trained on similar tasks tend to develop the same internal concepts and mechanisms, so what is learned about one model may carry over to others.