The Platonic Representation Hypothesis
2024/05/13 by Minyoung Huh, Huh, Minyoung, Brian Cheung +5 · 37 voices · 100 citations
Arts and Humanities · #Classical Philosophy and Thought
paper · pdf · doi:10.48550/arxiv.2405.07987
Abstract
We argue that representations in AI models, particularly deep networks, are converging. First, we survey many examples of convergence in the literature: over time and across multiple domains, the ways by which different neural networks represent data are becoming more aligned. Next, we demonstrate convergence across data modalities: as vision models and language models get larger, they measure distance between datapoints in a more and more alike way. We hypothesize that this convergence is driving toward a shared statistical model of reality, akin to Plato's concept of an ideal reality. We term such a representation the platonic representation and discuss several possible selective pressures toward it. Finally, we discuss the implications of these trends, their limitations, and counterexamples to our analysis.
Cited by
- SeeSE3: Emergence of 3D Space in Vision Features
- The Anatomy of a Truth Direction: Knowledge-Dependent Dimensionality, a Relational Law, and a Convergent Category Geometry in Small Language Models
- Capability from Access Structure, Not Scale: Lower Bounds and Pre-Registered Tests for Hybrid Sequence Models
- Attacking Graph Foundation Models Through Their Shared Representation
- Not-So-Strange Love: Language Models and Generative Linguistic Theories are More Compatible than They Appear
- Latent Communication Between Language Model Agents: Channels, Alignment, and the Limits of Text
- Language Game: Talking to Non-Human Systems
- Across the Levels of Analysis: Explaining Predictive Processing in Humans Requires More Than Machine-Estimated Probabilities
- Bootstrapping Life-Inspired Machine Intelligence: The Biological Route from Chemistry to Cognition and Creativity
- Linguists should learn to love speech-based deep learning models
- You can’t fight in here! This is BBS!
- Beyond Converging Representations A Philosophical Response on the Interpretation Risks of Scientific Foundation Models
- What does it mean to understand language?
- Predicting upcoming visual features during eye movements yields scene representations aligned with human visual cortex
- InputDSA: Demixing then Comparing Recurrent and Externally Driven Dynamics
- Words That Make Language Models Perceive
- Disentangling the Factors of Convergence between Brains and Computer Vision Models
- Representation biases: will we achieve complete understanding by analyzing representations?
- Harnessing the Universal Geometry of Embeddings
- Questioning Representational Optimism in Deep Learning: The Fractured Entangled Representation Hypothesis
- From superposition to sparse codes: interpretable representations in neural networks
- Shared Global and Local Geometry of Language Model Embeddings
- Representational Similarity via Interpretable Visual Concepts
- How linguistics learned to stop worrying and love the language models
- The Umwelt Representation Hypothesis: Rethinking Universality
- Emergence of Phonemic, Syntactic, and Semantic Representations in Artificial Neural Networks
- Reading Without a Reader: Large Language Models Collapse Reading and Writing into a Single Entangled Code
- Reference Feature Atlases for Mechanistic Auditing of Language Models
- Emergent Symbolic Structure in Health Foundation Models: Extraction, Alignment, and Cross-Modal Transfer
- Context Sensitivity Improves Human-Machine Visual Alignment
- Stop saying LLM: Large Discourse Models (LDM) and Artificial Discursive Agent (ADA)?
- In-Context Semi-Supervised Learning
- Large language models are not about natural language
- Unambiguous Representations in Neural Networks: An Information-Theoretic Approach to Intentionality
- Exploring possible vector systems for faster training of neural networks with preconfigured latent spaces
- Unleashing the Intrinsic Visual Representation Capability of Multimodal Large Language Models
- Unique Lives, Shared World: Learning from Single-Life Videos
- Network of Theseus (like the ship)
- TUNA: Taming Unified Visual Representations for Native Unified Multimodal Models
- Scaling and context steer LLMs along the same computational path as the human brain
- Multifractal Recalibration of Neural Networks for Medical Imaging Segmentation
- From Topology to Retrieval: Decoding Embedding Spaces with Unified Signatures
- Adversarial Confusion Attack: Disrupting Multimodal Large Language Models
- AI Consciousness and Existential Risk
- Better audio representations are more brain-like: linking model-brain alignment with performance in downstream auditory tasks
- PLATONT: Learning a Platonic Representation for Unified Network Tomography
- InstructMix2Mix: Consistent Sparse-View Editing Through Multi-View Model Personalization
- Tracing Multilingual Representations in LLMs with Cross-Layer Transcoders
- To Align or Not to Align: Strategic Multimodal Representation Alignment for Optimal Performance
- Time-Layer Adaptive Alignment for Speaker Similarity in Flow-Matching Based Zero-Shot TTS
- Mutual information and task-relevant latent dimensionality
- Rep2Text: Decoding Full Text from a Single LLM Token Representation
- Connecting the concepts of quantum state tomography and molecular representations for machine learning
- Superposition disentanglement of neural representations reveals hidden alignment
- Platonic Representations in the Human Brain: Unsupervised Recovery of Universal Geometry
- Back into Plato's Cave: Examining Cross-modal Representational Convergence at Scale
- RL makes MLLMs see better than SFT
- Visual Generation Unlocks Human-Like Reasoning through Multimodal World Models
- A Unified Geometric Space Bridging AI Models and the Human Brain
- Sprint: Sparse-Dense Residual Fusion for Efficient Diffusion Transformers
- VFM-VAE: Vision Foundation Models Can Be Good Tokenizers for Latent Diffusion Models
- Towards Error Centric Intelligence I, Beyond Observational Learning
- Position: Require Frontier AI Labs To Release Small "Analog" Models
- Rethinking the Simulation vs. Rendering Dichotomy: No Free Lunch in Spatial World Modelling
- Learning Model Representations Using Publicly Available Model Hubs
- Representational Alignment Across Model Layers and Brain Regions with Hierarchical Optimal Transport
- Contrastive Representation Regularization for Vision-Language-Action Models
- Uncovering the Computational Ingredients of Human-Like Representations in LLMs
- Understanding Adversarial Transfer: Why Representation-Space Attacks Fail Where Data-Space Attacks Succeed
- Beyond Token Probes: Hallucination Detection via Activation Tensors with ACT-ViT
- Convergence and Divergence of Language Models under Different Random Seeds
- Learning to See Before Seeing: Demystifying LLM Visual Priors from Language Pre-training
- Aligning Visual Foundation Encoders to Tokenizers for Diffusion Models
- Scalable GANs with Transformers
- Does Weak-to-strong Generalization Happen under Spurious Correlations?
- mini-vec2vec: Scaling Universal Geometry Alignment with Linear Transformations
- Toward a Theory of Generalizability in LLM Mechanistic Interpretability Research
- Guiding Evolution of Artificial Life Using Vision-Language Models
- Exposing Hallucinations To Suppress Them: VLMs Representation Editing With Generative Anchors
- Are Language Models Models?
- DS@GT ARC at ImageCLEFmedical 2026: Architectural Diversity for Concept Detection and Foundation-Model Scaling for Caption Prediction in Medical Image Analysis
- What Capable Agents Must Know: Selection Theorems for Robust Decision-Making under Uncertainty
- Knowledge Transfer from Interaction Learning
- \boldsymbolλ-Orthogonality Regularization for Compatible Representation Learning
- Charting trajectories of human thought using large language models
- MOCHA: Multi-modal Objects-aware Cross-arcHitecture Alignment
- Discovering Divergent Representations between Text-to-Image Models
- ALIGNS: Unlocking nomological networks in psychological measurement through a large language model
- Visual Representation Alignment for Multimodal Large Language Models
- RL's Razor: Why Online Reinforcement Learning Forgets Less
- What if I ask in alia lingua? Measuring Functional Similarity Across Languages
- Natural Latents: Latent Variables Stable Across Ontologies
- Prompt the Unseen: Evaluating Visual-Language Alignment Beyond Supervision
- Make me an Expert: Distilling from Generalist Black-Box Models into Specialized Models for Semantic Segmentation
- PiCSAR: Probabilistic Confidence Selection And Ranking for Reasoning Chains
- Learning from nature: insights into GraphDOP's representations of the Earth System
- Large Language Models Show Signs of Alignment with Human Neurocognition During Abstract Reasoning
- Cross-Model Semantics in Representation Learning
- CauKer: Classification Time Series Foundation Models Can Be Pretrained on Synthetic Data
- Text-Attributed Graph Anomaly Detection via Multi-Scale Cross- and Uni-Modal Contrastive Learning
Discussions
- Mutlimodal neural networks converge to a shared statistical model of reality [hn, 34 points, 6 comments]
- 2/ Ex 1: Platonic representations Whether you train on vision or langauge, big nets have aligned representations with similar structures arxiv.org/abs/2405.07987 So, networks may be extracting some [bsky, 23 points, 3 comments]
- Last, a talk on "The Platonic Representation Hypothesis" (arxiv.org/abs/2405.07987) at UniReps. This talk says: don't worry, we don't need to equip NNs with special distance measures, these structure [bsky, 20 points, 1 comments]
- Some people have even proposed that this is inevitable, not because networks are overfitting to the data distribution, but because all networks will at some point, given enough data & enough parametri [bsky, 13 points, 1 comments]
- This is from 2024 but pretty fascinating: "We argue that representations in AI models, particularly deep networks, are converging" arxiv.org/abs/2405.07987 It's an interesting question whether there i [bsky, 11 points, 3 comments]
- Part of why they're drawing that bigger claim is because of an AI concept called the Platonic Representation Hypothesis, which is a cute metaphor, but isn't the same as Platonic Idealism being proven [bsky, 10 points, 1 comments]
- 2/3 Studie arxiv.org/abs/2405.07987 [bsky, 8 points, 1 comments]
- Have you read the Platonic Representation Hypothesis paper? Essentially argues that models across modalities converge with enough data arxiv.org/abs/2405.07987 [bsky, 6 points, 0 comments]
- it turns out (mostly) no! arxiv.org/abs/2405.07987 [bsky, 6 points, 1 comments]
- Vision models (backbones & foundation models alike) seem to learn transferable features that are relevant across many tasks. Recent work even suggests we are converging towards the same "Platonic" rep [bsky, 5 points, 1 comments]
- Un argument pour : arxiv.org/abs/2405.07987 Un argument contre : arxiv.org/abs/2604.18572 Les deux sont très bien. Le débat fait rage au sein de mon labo (et même entre mes doctorant) et donne lieu pa [bsky, 5 points, 2 comments]
- When I first read this paper, I instinctively scoffed at the idea. But the more I look at empirical results, the more I’m convinced this paper highlights something fundamentally amazing. Lots of excit [bsky, 3 points, 3 comments]
- This one outlines the original idea: arxiv.org/abs/2405.07987. Really makes me more confident that there are real patterns and ML systems can track them to a certain extent. Of course industry practic [bsky, 3 points, 2 comments]
- Related, maybe [bsky, 3 points, 1 comments]
- The Platonic Representation Hypothesis [hn, 2 points, 1 comments]
- There is good science backing this up tbh, unlike intelligent design arxiv.org/abs/2405.07987 [bsky, 2 points, 2 comments]
- You’ve seen this yes? arxiv.org/pdf/2405.07987 [bsky, 2 points, 1 comments]
- Do you have any thoughts on how this relates to e.g. the Platonic Representation Hypothesis paper? arxiv.org/abs/2405.07987 [bsky, 2 points, 1 comments]
- related [bsky, 2 points, 0 comments]
- With the caveat that importance is best judged in hindsight, I quite liked "The Platonic Representation Hypothesis" arxiv.org/abs/2405.07987 [bsky, 2 points, 0 comments]
- Of course it doesn't "see" like us or "read" like us, This paper suggests that multimodality can be useful arxiv.org/pdf/2405.07987. Humans don't have an inherent concept of physics either, we gain an [bsky, 2 points, 1 comments]
- Great paper: "The Platonic Representation Hypothesis" Neural networks, trained with different objectives on different data and modalities, are converging to a shared statistical model of reality in th [bsky, 2 points, 0 comments]
- There a lot of different parallel explorations, and haven’t seen anyone attempt to link this phenomenon back to a broader framing like you put forward. But some come pretty close eg arxiv.org/abs/2405 [bsky, 1 points, 1 comments]
- Thanks to @sjvn.bsky.social for this article. Also thanks for pointing out some of the BS on arxiv lately. Here's another one talking about 'convergence' and ' statistical model of reality'. [bsky, 1 points, 1 comments]
- Then why are you posting? Get out of the cave, dummy! [bsky, 1 points, 0 comments]
- Hehe while I’m flattered you remember our conversation—and while, as an idealist, these notions have long been the scaffold of my thinking and that of many in mathematics—when folks online say “Platon [bsky, 1 points, 1 comments]
- ➕ Bonus: Theory can explain the “Platonic Representation Hypothesis”—the striking observation that different models often learn the same representations. arxiv.org/abs/2405.07987 With the right assump [bsky, 0 points, 0 comments]
- Read the 'Platonic representation hypothesis' paper and can't shake my suspicions (it's mainly about vision and doesn't mention audio then makes a lot of caveats), call me a cynic... arxiv.org/abs/240 [bsky, 0 points, 0 comments]
- "The Platonic Representation Hypothesis" 🤯 arxiv.org/abs/2405.07987 [bsky, 0 points, 0 comments]
- [2405.07987] The Platonic Representation Hypothesis — A thoughtful arXiv paper arguing for a “Platonic” view of representations—useful if you’re thinking about what learned features really are across [bsky, 0 points, 0 comments]
- There’s widespread suspicion that as model and data scale increases, model representations converge (e.g., the platonic representation hypothesis (arxiv.org/abs/2405.07987), in part supported by diffe [bsky, 0 points, 1 comments]
- Look forward to reading the paper. arxiv.org/pdf/2405.07987 [bsky, 0 points, 0 comments]
- On a more serious note, what implications does this result have wrt ontology? arxiv.org/abs/2405.07987 [bsky, 0 points, 0 comments]
- "representations in AI models, particularly deep networks, are converging. (...) this convergence is driving toward a shared statistical model of reality, akin to Plato's concept of an ideal reality." [bsky, 0 points, 1 comments]
- I think so! I am a believer in the platonic representation hypothesis, for models and biological brains :) arxiv.org/abs/2405.07987 [bsky, 0 points, 1 comments]
- The Platonic Representation Hypothesis [bsky, 0 points, 0 comments]
- It feels like you are almost referring to this and you might know, but since it was not linked in the post arxiv.org/abs/2405.07987 [bsky, 0 points, 0 comments]
Related