Revisiting the Platonic Representation Hypothesis: An Aristotelian View
Fabian Gröger, Shuo Wen, Maria Brbić
2026-02-18
Summary
This paper investigates whether different neural networks, even those trained on different types of data, are learning to represent the world in the same way, an idea called the Platonic Representation Hypothesis.
What's the problem?
Previous methods for comparing how different neural networks 'see' things were flawed because larger networks always seemed more similar, regardless of whether they actually were. This made it hard to tell if networks were truly converging on a shared understanding of reality or if the similarity was just an artifact of their size. Basically, the tools used to measure similarity were giving misleading results.
What's the solution?
The researchers developed a new way to compare network representations that corrects for the size of the network. They used a statistical technique called 'permutation-based null-calibration' which essentially removes the bias caused by network depth and width, giving a more accurate measure of true similarity. They then reapplied this new method to the question of whether networks are learning similar representations.
Why it matters?
The findings suggest that networks aren't necessarily learning a single, universal model of reality as previously thought. Instead, they seem to be agreeing on how things relate to each other locally – meaning they understand which features tend to appear together – but don't necessarily share a complete, global understanding. This leads the authors to propose a new idea, the Aristotelian Representation Hypothesis, where networks converge on shared local relationships rather than a single 'Platonic' ideal.
Abstract
The Platonic Representation Hypothesis suggests that representations from neural networks are converging to a common statistical model of reality. We show that the existing metrics used to measure representational similarity are confounded by network scale: increasing model depth or width can systematically inflate representational similarity scores. To correct these effects, we introduce a permutation-based null-calibration framework that transforms any representational similarity metric into a calibrated score with statistical guarantees. We revisit the Platonic Representation Hypothesis with our calibration framework, which reveals a nuanced picture: the apparent convergence reported by global spectral measures largely disappears after calibration, while local neighborhood similarity, but not local distances, retains significant agreement across different modalities. Based on these findings, we propose the Aristotelian Representation Hypothesis: representations in neural networks are converging to shared local neighborhood relationships.