{"id":8,"url":"https://pm.philipcastiglione.com/papers/8.json","title":"The Platonic Representation Hypothesis","read":true,"authors":"Minyoung Huh, Brian Cheung, Tongzhou Wang, Phillip Isola","year":2024,"auto_summary":"This paper introduces the Platonic Representation Hypothesis, which posits that as AI models, particularly deep neural networks, grow in complexity and size, they are converging towards a shared statistical model of reality. This convergence is akin to Plato's concept of an ideal reality, where models across different modalities (e.g., vision and language) begin to represent data in increasingly similar ways.\n\nKey points from the paper include:\n\n1. **Convergence Evidence**: The authors survey literature showing that neural networks trained on different tasks and data modalities are converging in their representational spaces. This convergence is observed across various domains and over time.\n\n2. **Platonic Representation**: The hypothesis suggests that this convergence is towards a \"platonic representation,\" a shared statistical model of the underlying reality that all these models are approximating.\n\n3. **Selective Pressures**: Several factors drive this convergence, including the scaling of model size, diversity of data and tasks, and the inherent biases in deep learning models towards simpler solutions.\n\n4. **Implications**: The convergence has several implications, such as the potential for training data to be shared across modalities, ease of translation and adaptation between modalities, and possible reductions in model hallucinations and biases as models scale.\n\n5. **Limitations and Counterexamples**: The paper acknowledges that different modalities may capture different information, which could limit convergence. Additionally, not all representations may converge, especially in specialized domains with unique tasks.\n\n6. **Experiments and Metrics**: The authors conduct experiments to measure alignment between models using various metrics, such as mutual nearest-neighbor and CKA, and find that alignment increases with model competence and scale.\n\n7. **Contrastive Learning**: The paper discusses how certain contrastive learning objectives naturally lead to representations that approximate pointwise mutual information (PMI), which is a measure of statistical dependency between variables.\n\nOverall, the paper provides a theoretical framework and empirical evidence for the idea that AI models are converging towards a unified representation of reality, driven by task generality, model capacity, and simplicity biases.","notes":{"id":8,"name":"notes","body":"\u003ch1\u003eInformation\u003c/h1\u003e\u003cdiv\u003e\u003cbr\u003eGenerally this covers for representations that are vector embeddings.\u003cbr\u003e\u003cbr\u003e\u003c/div\u003e\u003cdiv\u003e\u003cbr\u003eDifferent perceptions exist. Many, overlap (vision and touch can both describe the shape of something, though only vision describes its colour).\u003cbr\u003e\u003cbr\u003e\u003c/div\u003e\u003cdiv\u003e\u003cbr\u003ePerceptions (eg from different kinds of sensors) can be sufficiently different that representations might not converge in unimodal (or non-overlapping) systems.\u003cbr\u003e\u003cbr\u003e\u003c/div\u003e\u003cdiv\u003e\u003cbr\u003eConvergence correlates with increasing performance (which increases with scale) - makes sense as you need a good representation to get good outcomes.\u003cbr\u003e\u003cbr\u003e\u003c/div\u003e\u003cdiv\u003e\u003cbr\u003eEarly layers contain less abstract concepts and are more interchangeable. Primitives are easier to derive the platonic representation of. (Rods and cones, colours, vs full knowledge of an obscure pop culture figurine, its history etc).\u003cbr\u003e\u003cbr\u003e\u003c/div\u003e\u003cdiv\u003e\u003cbr\u003eRepresentations converge across modalities (where there is overlap).\u003cbr\u003e\u003cbr\u003e\u003c/div\u003e\u003cdiv\u003e\u003cbr\u003eModel are increasingly aligning to brains (there’s a paper I plan to read about CNNs being aligned with the visual cortex de novo). This is because they’re solving the same problem.\u003cbr\u003e\u003cbr\u003e\u003c/div\u003e\u003cdiv\u003e\u003cbr\u003eConvergence is proposed to occur via 3 pathways:\u003cbr\u003e\u003cbr\u003e\u003c/div\u003e\u003col\u003e\u003cli\u003eTask generality - each training datapoint and task places a constraint on the model. fewer representations exist which solve more constrained models. generality emerges as the viable solution.\u003c/li\u003e\u003cli\u003eModel capacity - bigger models are more likely to converge to a shared representation, because they each are more likely to overlap over the ideal representation.\u003c/li\u003e\u003cli\u003eSimplicity bias - since the bigger a deep network is, the more likely it is to find simple fits, this should also lead to convergence towards simple solutions together.\u003c/li\u003e\u003c/ol\u003e\u003cdiv\u003e\u003cbr\u003eImplications:\u003cbr\u003e\u003cbr\u003e\u003c/div\u003e\u003cul\u003e\u003cli\u003escaling is sufficient but not necessarily efficient\u003c/li\u003e\u003cli\u003etraining data can (should? for better sample efficiency) be shared across modalities\u003c/li\u003e\u003cli\u003etransitioning across modalities should be easy (to the extent that they overlap, eg. like language and vision)\u003c/li\u003e\u003cli\u003escaling might reduce hallucinations and bias (by amplifying it less)\u003c/li\u003e\u003c/ul\u003e\u003cdiv\u003e\u003cbr\u003eSpecial purpose intelligences converge less (these can get you narrow wins)\u003cbr\u003e\u003cbr\u003e\u003c/div\u003e\u003ch1\u003eQuestions\u003c/h1\u003e\u003cdiv\u003e\u003cbr\u003eContra convergence via task generality - what about overfitting?\u003cbr\u003e\u003cbr\u003e\u003c/div\u003e\u003ch1\u003eTakeaways\u003c/h1\u003e\u003cdiv\u003e\u003cbr\u003eFundamentally; there is a true (platonic) state of the world; perception gives us a view into that state; AI systems that build representations of the world based on perception data (eg. deep neural networks) converge towards that true state as they grow.\u003cbr\u003e\u003cbr\u003e\u003c/div\u003e\u003cdiv\u003e\u003cbr\u003eLLMs (transformer models driven by SGD generally?): the ultimate in offline development of a useful abstract representation of the input data. What they’re achieving (in this step) is building a good representation.\u003cbr\u003e\u003cbr\u003e\u003c/div\u003e\u003cdiv\u003e\u003cbr\u003e\u003c/div\u003e","record_type":"Paper","record_id":8,"created_at":"2024-12-10T03:55:11.817Z","updated_at":"2024-12-10T04:00:31.345Z"},"created_at":"2024-12-10T03:54:17.998Z","updated_at":"2024-12-10T04:00:31.346Z"}