PaperlayerView exact revision on arXiv

Source-linked AI summary

The Platonic Representation Hypothesis

Minyoung Huh, Brian Cheung, Tongzhou Wang, Phillip Isola

arXiv:2405.07987v5cs.LGcs.AIcs.CVcs.NE

TL;DR

The paper asks whether increasingly aligned representations across neural networks and modalities reflect convergence toward a common model of reality. It synthesizes prior evidence, evaluates representational alignment across models, and argues that this convergence is associated with shared structure and downstream performance, while acknowledging important limits from information loss and observation constraints.

  • Problem

    The paper addresses the question of whether representations learned by models with different architectures, objectives, data, and modalities are becoming aligned, and where this convergence ends.

  • Method

    The paper surveys evidence for representational convergence and measures alignment using kernel-based representations and a mutual nearest-neighbor metric.

  • Results

    The paper finds convergence across models and modalities: stronger language-modeling performance tends to align with vision models, while alignment improves alongside commonsense reasoning and mathematical problem-solving performance.

  • Takeaways & Limitations

    The findings support interpreting convergence as movement toward a shared statistical model of the underlying reality that generates observations.

  • Takeaways & Limitations

    The argument strictly applies only to bijective projections, while lossy or stochastic observations can limit convergence because different sensors may capture different information.

Abstract

from arXiv · show

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.

1. Introduction

The paper argues that neural networks trained with different objectives, data, architectures, and modalities are increasingly converging on aligned representations. It hypothesizes that this convergence may approach a shared statistical representation of the underlying reality that generates observed data.

  • Neural networks trained on different objectives, data, and modalities are converging toward a shared statistical model of reality.
  • The paper focuses on representational convergence: increasing similarity in how different neural network models represent datapoints.
  • The authors frame convergence as an open question concerning its causes, continuation, and endpoint.
  • The central hypothesis is that models seek representations of the joint distribution over world events that generate observed data.
  • The proposed converged representation is called the “platonic representation,” echoing Plato’s idea of an ideal reality underlying sensory experience.

2. Representations are converging

Evidence across architectures, objectives, datasets, and modalities indicates that neural-network representations are becoming increasingly aligned. The paper measures this convergence through kernel-based similarity and relates stronger alignment to model competence and downstream performance.

  • Representational alignment compares the similarity structures induced by different vector-embedding representations.The paper represents each model through its kernel and compares kernels using alignment metrics.
  • Prior work finds compatible representations across image datasets, text models, languages, individual neurons, and model scales.Model stitching, zero-shot transfer, Rosetta Neurons, and scaling studies provide distinct examples of convergence.
  • The paper evaluates 78 vision models and finds that higher VTAB transfer performance corresponds to more tightly clustered representations.Models vary in architecture, training objective, and dataset, while alignment is measured within performance bins.
  • Cross-modal alignment increases with language-modeling ability: stronger language models tend to align more with stronger vision models.The analysis uses paired images and captions from the Wikipedia captions dataset and mutual nearest-neighbor alignment.
  • Closer alignment to vision models is associated with better Hellaswag commonsense reasoning and GSM8K mathematical problem-solving performance.Hellaswag shows a linear relationship with alignment, whereas GSM8K exhibits an emergence-like trend.

3. Why are representations converging?

The paper proposes that increasing data, task generality, model capacity, and simplicity bias progressively narrow the set of viable representations, promoting convergence. These pressures may make larger models converge toward shared solutions.

  • Convergence via Task Generality: As data and tasks scale, additional constraints shrink the volume of representations that satisfy them.The paper frames multitask learning as reducing the number of competent solutions as models solve more tasks simultaneously.
  • Convergence via Task Generality: There are fewer representations competent for N tasks than for M < N tasks, narrowing the possible solution set as task coverage expands.This is the paper’s Multitask Scaling Hypothesis.
  • Convergence via Task Generality: As data scale, empirical-risk optimization improves population-risk performance and better captures statistical structure in the true data-generating process.The paper links larger datasets to progressively tighter constraints on representations that fit reality-relevant structure.
  • Convergence via Model Capacity: Bigger models are more likely to converge to a shared representation than smaller models when larger function classes can cover an optimum.With the same objective, sufficiently large models may find similar minimizers even across different architectures.
  • Convergence via Model Capacity: Deep networks favor simple fits, and this simplicity bias strengthens with model size, reducing the solution space toward which models converge.The bias may arise from explicit regularization or implicit preferences for simpler solutions.

4. What representation are we converging to?

The paper hypothesizes that representational convergence approaches a statistical model of the underlying reality generating observations. In an idealized multimodal setting, contrastive learners recover a shared pairwise-statistics kernel, with color cooccurrence providing an empirical illustration.

  • What representation are we converging to?: The proposed endpoint of convergence is a statistical model of the underlying reality that generates observations and supports many reality-grounded tasks.The paper calls this hypothetical converged representation the “platonic representation.”
  • An idealized world: The formalization treats world events as samples from an unknown distribution P(Z), observed through bijective deterministic mappings into modalities such as pixels, sounds, and words.The paper presents this as one concrete candidate form for the platonic representation, not the only possible formalization.
  • A family of contrastive learners converge to a representation of P(Z): Under sufficient data and optimization, contrastive learners converge to representations whose kernel equals the pointwise mutual information kernel up to a constant offset.The result applies under the paper’s mild smoothness conditions.
  • A family of contrastive learners converge to a representation of P(Z): For any bijective discrete modality in the idealized world, representational convergence reaches the same kernel, encoding pairwise statistics of P(Z).The modality-agnostic formulation is intended to explain convergence across observation types.
  • An idealized world: Color representations learned from image and language cooccurrences roughly recover the same perceptual organization as CIELAB color space.The paper reports this as a real-data case study and notes increasing similarity as language models scale.
  • What representation are we converging to?: The resulting framework offers a path toward understanding a unified representation that is proficient across domains and modalities and grounded in statistical properties of the underlying world.The authors explicitly state that Section 6 discusses limitations of the simple model.

5. What are the implications of convergence?

The paper argues that representational convergence could make scaling useful across modalities, improve transfer between domains, and potentially reduce hallucination and some bias, though efficiency still varies.

  • Scaling is sufficient, but not necessarily efficient: Scaling resources are associated with converging representations across modeling choices and modalities, but methods can scale with different efficiency.Successful methods must also satisfy general requirements such as consistent estimation and modeling pairwise statistics of P(Z).
  • Training data can be shared across modalities: A modality-agnostic representation implies that image and language data could jointly improve representation learning in both vision and language models.The paper notes that image training has already improved text performance and suggests cross-modal conversion ratios may exist.
  • Ease of translation and adaptation across modalities: Aligned representations should make transitions between modalities easier and may support conditional generation, representation-conditioning, and mapping between domains without paired data.The paper connects this possibility to the observed ease of conditional generation and recent representation-conditioning results.
  • Ease of translation and adaptation across modalities: Cross-modal data may primarily provide sample efficiency when different modalities contribute to a shared representation.The paper frames this as a consequence of sharing information across modalities rather than as a separate modeling mechanism.
  • Scaling may reduce hallucination and bias: If scaling drives models toward an accurate reality model, hallucinations may decrease and bias may more closely reflect training-data bias rather than amplify it.This implication depends on future training data being sufficiently lossless and diverse, and does not imply that bias disappears.

6. Counterexamples and limitations

The convergence hypothesis is limited by modality-specific information, non-bijective observations, incomplete convergence across domains, sociological influences, task-specific shortcuts, measurement choices, and unresolved alignment strength.

  • Different modalities may contain different information: Different modalities may contain unique information, preventing models with fundamentally different observations from converging to identical representations.The paper gives touch-versus-vision and abstract-language examples, and explicitly notes that sensors may capture different information.
  • Different modalities may contain different information: The mathematical argument strictly applies only to bijective projections of Z, not lossy or stochastic observations, so a nuanced hypothesis is needed for non-bijective inputs and abstract concepts.The paper proposes that lower-information representations may align with higher-information ones only up to a limited level.
  • Different modalities may contain different information: More descriptive captions produce better language-vision alignment in a preliminary test, consistent with higher-information inputs supporting stronger cross-modal correspondence.The reported experiment uses the DCI dataset and compares caption densities generated or summarized with LLaMA3-8B-Instruct.
  • Not all representations are presently converging: Convergence is not yet equally evident across domains: robotics lacks a standardized world-state representation and faces expensive, slow hardware that limits training-data quantity and diversity.The paper’s main argument focuses on vision and language, while expecting other modalities to follow similar trends remains prospective.
  • Sociological bias and special-purpose intelligences: Researcher preferences may steer models toward human-like representations, while special-purpose intelligences may use efficient task-specific shortcuts detached from reality.The latter boundary becomes more relevant if energy and compute constrain continued scaling.
  • How do we measure alignment?: The merits and deficiencies of alignment measures remain debated, and the observed rise to 0.16 on the mutual nearest-neighbor metric leaves its strength uncertain against a theoretical maximum of 1.The paper discusses alternative metrics in Appendix A and leaves the interpretation of the remaining alignment gap open.

A. Mutual k-Nearest Neighbor Alignment Metric

The paper defines alignment through similarity between representation-induced kernels and introduces local kernel alignment metrics that connect CKA with mutual nearest-neighbor overlap.

  • Mutual nearest-neighbor metric: The mutual k-nearest-neighbor metric measures the average overlap between the two models’ respective nearest-neighbor sets.The intersection size is used to quantify the overlap for each feature pair.
  • Mutual nearest-neighbor metric: For cross-modal samples, the method computes feature pairs from two models and forms each feature’s nearest-neighbor sets before comparing those sets.For single-domain measurements, corresponding samples are treated as identical inputs.
  • Metric choice and experiment: The authors chose nearest-neighbor alignment because CKA showed weak trends and imposes a strict notion of similarity that may be unnecessary for unrelated items.They compare language-vision alignment across k values and report more coherent trends as k decreases.
  • From CKA to local alignment: CKA measures global congruence between kernels, with alignment ranging from 0 to 1, whereas restricting cross-covariance to nearest neighbors emphasizes local alignment.CKA is invariant to isotropic scaling and compares similarity structure across all samples.
  • From CKA to local alignment: The CKNNA weighting selects pairs that are mutual nearest neighbors, and as k approaches the kernel dimension it recovers ordinary CKA.Relaxing equal-neighbor distances leads back to the mutual nearest-neighbor metric.
  • Metric choice and experiment: In Figure 10, k = 1024 recovers CKA, while smaller k values closely resemble mutual nearest-neighbor alignment and make the trend more pronounced.Each line averages alignment scores across the evaluated language models for a specific k.

B. Consistency across various metrics

The paper compares alignment metrics across vision models and across modalities, finding that metric choice and task affect cross-modal alignment while most vision metrics are highly correlated.

  • Metric properties: The metric comparison includes symmetric, global, ordinal, and batchable properties.Global metrics use all samples when computing distances, whereas ordinal metrics preserve distance ordering.
  • Vision-vision comparison: Most alignment metrics are highly correlated when evaluated across 78 vision models.Figure 12 reports Spearman rank correlations across metrics and batch sizes, with all p-values below 2.24 × 10−105.
  • Cross-modal comparison: Cross-modal alignment sensitivity depends on both the metric used and the tasks on which vision models were trained.The paper visualizes these comparisons in Figures 13 and 14.
  • Vision-vision comparison: The paper’s vision-vision analysis uses mutual k-NN with k = 10 and batch size 1000.Figure 12 identifies this as the first metric configuration used for the main vision-vision analysis.

C. Experiments on Evaluating Alignment and Convergence

The experiments evaluate representational alignment across diverse vision and language models using nearest-neighbor structure, while testing representation quality and cross-modal layer alignment.

  • Experimental setup: The study measures representational alignment across off-the-shelf models spanning multiple scales and modalities.The vision set contains 78 models with varied architectures, training objectives, and datasets.
  • Vision-vision alignment and representation quality: Representation quality is evaluated by linear probing on all 19 VTAB classification tasks.A representation is considered to solve a task when it reaches at least 80% of the best performance among the 78 models.
  • Vision-vision alignment and representation quality: The vision-vision alignment metric uses k = 10 nearest neighbors over 1000 Places-365 validation images disjoint from VTAB datasets.The shared natural-image content of Places-365 and VTAB is noted despite the dataset split.
  • Cross-modal alignment: Cross-modal comparisons pair image representations with caption representations from BLOOM, OpenLLaMA, LLaMA, and newer language-model families.The main vision models include ViTs trained with classification, MAE, DINOv2, and CLIP objectives.
  • Cross-modal alignment: Cross-modal alignment uses 1024 WIT samples, compares pairwise layers, and takes the maximum alignment score.Vision uses class tokens, while language uses average-pooled layer representations.
  • Cross-modal alignment: Prompting produces trends similar to average pooling but slightly lower alignment scores.

D. Color Cooccurrence Experiment

The color experiment constructs comparable representations from perceptual color space, visual cooccurrence, and language, using dissimilarities and multidimensional scaling.

  • Color representations: The experiment visualizes four color representations using multidimensional scaling from pairwise dissimilarities.The best of 1,000 fits is selected and then visually aligned with CIELAB space.
  • Perceptual representation: CIELAB provides a perceptually uniform representation in which numerical changes correspond to similar perceived color changes.
  • Vision cooccurrence: The vision representation estimates color cooccurrence from 300,000 nearby pixel-color pairs in CIFAR-10.Colors are quantized in RGB space, and normalized counts are used to compute an empirical pointwise mutual information matrix.
  • Language cooccurrence: The language representation uses 20 color-word pairs derived from participant naming data and assigns each word a CIELAB centroid.Words occurring fewer than 100 times are filtered, producing a slightly different set from the 18 colors reported by Abdou et al. (2021).
  • Language cooccurrence: The language inputs include templated sentences such as “This color is <col>.”
  • Method choice: The method avoids linear regression to CIELAB because it distorts distances and can overfit with only 20 samples.Instead, multidimensional scaling is used to preserve distances.
  • Language embeddings: SimCSE and RoBERTa embeddings provide alternative language-based dissimilarity matrices from Euclidean distances.RoBERTa uses a 4096-dimensional embedding formed from the last four layers.

E. Caption Density Experiments

The caption-density experiment generates captions at different lengths and tests whether denser descriptions align more strongly with vision representations.

  • Caption generation: LLaMA3-8B-Instruct generates captions of varying densities for images from the Densely Captioned Images training split.
  • Alignment result: Alignment scores improve as generated caption length increases.This result is reported in Figure 9 as support for testing denser captions.

F. Analysis of Contrastive Learners

The analysis shows that both binary NCE and InfoNCE contrastive objectives recover PMI-based scores, while exact inner-product representation requires a positive-semidefinite shifted KPMI kernel. Under sufficiently smooth distributions or high sampling rates, this representation can exist, although the stated smoothness condition is restrictive.

  • Contrastive objectives: Both binary NCE and InfoNCE objectives are minimized by pointwise mutual information (PMI).The analysis considers bivariate scores g and optimizes them within an inner-product representation family.
  • Contrastive objectives: Positive pairs come from the coordinated distribution, whereas negative pairs are sampled independently from its marginals.
  • Contrastive objectives: For general τ ≠ 1, the learned score and corresponding representation recover KPMI up to an offset and scale.
  • Representability conditions: An exact representation ⟨fX(xa), fX(xb)⟩ = KPMI(xa, xb) + C exists when the shifted KPMI kernel is positive semi-definite.This criterion follows from eigendecomposition and governs whether KPMI + C can be represented as inner products.
  • Representability conditions: Proposition F.1 establishes positive semidefiniteness for some C when off-diagonal KPMI values are bounded and the joint distribution is sufficiently smooth.The proof uses non-positive off-diagonal entries and diagonal dominance after choosing C appropriately.
  • Representability conditions: Smoothness, continuity, or sufficiently high sampling rates can make the PMI kernel exactly representable, but the proposition’s condition is somewhat strict.The remark gives linearly decaying, sufficiently similar off-diagonal terms as an example satisfying the condition.
Loading 2405.07987v5…