More Data Cannot Break a Symmetry: Identifiability by Design

Jing Xu; Christopher Kanan

cs.LGAug 27, 2026

More Data Cannot Break a Symmetry: Identifiability by Design

Authors: Jing Xu, Christopher Kanan

Organizations: Department of Computer Science, University of Rochester

Abstract

Unsupervised representational alignment recovers a stimulus-by-stimulus correspondence from geometry alone, but the automorphism group of the stimulus geometry bounds what any such alignment can identify, before data exist. The obvious diagnostic for this degeneracy, the cheapest non-identity relabelling, ranks two published designs in the wrong order, because dense sampling creates near-duplicates whose transposition is nearly free. We turn this known invariance (Demetci et al., 2024) into a design-time diagnostic and intervention. In colour, where candidate geometries have closed form, we show that the failure is structural: sixty-four times the restart budget leaves a symmetric design unmoved while an asymmetric set at the same N recovers every time. Discriminating representational models and recovering a correspondence are essentially uncorrelated objectives (r = -0.02 over 3,000 subsets). Choosing nine colours by this diagnostic alone, without consulting any learned representation, moves all 93 model representations away from the degenerate point and cuts catastrophic alignment failures from 75% to 2% with the models, the layers, N and the solver all held fixed. The same risk arises wherever a regular design meets its candidate geometry's isometry group, including evenly spaced orientations, tones, or motion directions, and the check costs one function call before data collection.

Figures & tables

Figure 1: Under the HSV cylinder every rotation of the nine-colour set costs within

0.003

of the identity (the nominal design is exactly equally spaced; using the calibrated sRGB values supplied with the dataset introduces deviations of at most

0.003

); under the observers’ own judgements (rotation cost

0.26

) or under nine colours drawn from the colour solid, it does not. Lower panels: the two correspondences a zero cost cannot distinguish from the truth.

Figure 2: Computation rescues an optimisation problem and cannot rescue an identifiability one. The ring is unmoved by sixty-four times the search; the 93-colour displacement falls to

0.052

while exact accuracy stays at

0\%

, which is why displacement is the primary outcome.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Design	Field	Group	Catastrophic cost
Nine saturated hues	colour	$D_{9}$	$0.000$ (exact)
Twelve orientations	V1	$D_{12}$	$0.000$ (exact)
Twelve motion directions	MT	$D_{12}$	$0.000$ (exact)
Twelve tones within an octave	audition	$D_{12}$	$0.000$ (exact)
$5\times 5$ factorial grid	psychophysics	$D_{4}$	$0.000$ (exact)
THINGS objects ( $n=200$ )	object vision	—	$0.224$

Appendix

Table 1: Catastrophic cost across subfields. Zeros marked exact are achieved by the design’s own symmetry group and so are not search bounds. Each is conditional on its candidate geometry exactly as the colour result is: the orientation zero assumes orientation is an angle on a circle, the standard model rather than a fact. THINGS uses human similarity embeddings from the SPoSE model ( Hebart et al., 2020 ) .

Figure 3: Catastrophic cost against set size for three sampling schemes. Keeping stimuli on the hue circle at

S=V=1

(orange) leaves the cost at zero regardless of

N

; any departure from the invariant submanifold lifts it immediately. The published 93-colour set (star, upper right) falls on the curve at its own

N

: it is not special, just not on the circle. The matched-

N

contrast at

N=9

rules out the reading that the 93-colour success is simply more data. Bands show 10th–90th percentile over 40 random draws.

Figure 4: Design-time prescription. Before data collection, compute the catastrophic cost of the stimulus set under every candidate geometry and take the worst case

\Gamma(S)

. A zero means the correspondence cannot be identified under that geometry; the remedy is to vary the sampling away from the invariant submanifold, not to add more stimuli along it.

Figure 5: Four designs matched at nine stimuli, shown to the same 93 representations. Left: the closest any layer of each model comes to zero, models ranked within each design. The correspondence-optimised set moves every model further out, though the spread grows in proportion to the shift rather than beyond it (SD over median

0.30

against

0.29

for the ring), so the models are not made more separable. The geometry-discrimination set tracks the random draw almost exactly. Right: how often the recovered correspondence is catastrophic, with the design-time cost beside each bar.

Figure 6: Catastrophic cost across 93 model representations. Left: the published nine-colour set against 930 matched-size draws from the colour solid, the same representations and the same

N

in both cases. No representation sits at the degenerate point, and the solid draws sit further from it. Right: the closest any layer of each model comes to zero. The line marks the human observers’ catastrophic cost (

0.085

); five models have at least one layer below it.

Representation Identifiability

Representation identifiability studies when learned internal codes can be uniquely recovered from observations or behavior, and which ambiguities remain.

See all 41 papers

May 22, 2026

Representation Alignment Rests on Linear Structure

We investigate the Platonic Representation Hypothesis (PRH) through a tripartite statistical framework of representations: signal, bias, and noise. {1) Signal:} We propose that Platonic alignment arises from the universal relationship between objects and attributes, which is encoded linearly in representations according to the Linear Representation Hypothesis (LRH). We provide evidence that LRH helps explain PRH by extracting linear object-attribute features with sparse autoencoders and showing that these sparse representations often exhibit stronger cross-modal alignment than their dense counterparts. {2) Bias:} Models have different implicit biases due to the diverse architectures and training procedures used. We show that this difference can be partially mitigated. Centering and normalization consistently improve cross-model alignment. {3) Noise:} Finite-sample training leads to noise in representations. We provide evidence that representational noise is driven by data scarcity by revealing a strong and consistent positive correlation between word frequency and alignment in LLMs and text embedding models. Synthesizing signal, bias, and noise, we propose a statistical model that refines the Linear Representation Hypothesis and explains further phenomena related to the alignment of representations emerging from diverse modern AI architectures.

Kiril Bangachev, Guy Bresler, Yury Polyanskiy

Department of Electrical Engineering and Computer Science · Massachusetts Institute of Technology

Linear Representation Hypothesis Cross-Modal Alignment Representation Alignment Platonic Representation Hypothesis

Ask PDF

May 20, 2026

Stimulus symmetries can confound representational similarity analyses

What can representational similarity matrices (RSMs) tell us about a neural code? As the popularity of these summary statistics grows, so too does the need for a more complete characterization of their properties. Here, we show that symmetries in network inputs can confound RSM-based analyses. Stimulus symmetries render many representations functionally equivalent, but these different configurations can lead to different RSMs. These different RSMs reflect qualitatively different representational geometries. We show that stochastic gradient descent or energetic regularization can generate sparse, drifting codes, leading in turn to drifting RSMs. Moreover, we demonstrate that these phenomena are present in networks trained to encode image data, where the symmetry is latent. Our results illustrate the challenges inherent in comparing nonlinear neural codes, when functionally-equivalent representations are not related by a simple rotation.

Farhad Pashakhanloo, Jacob A. Zavatone-Veth

Center for Brain Science and 2Society of Fellows Harvard University, Cambridge, MA, USA

Representational Similarity Analysis Representation Learning Neural Representation Geometry Sparse Neural Networks Neural Encoding

Ask PDF

Aug 18, 2026

Bidirectional representational alignment between biological and artificial neural networks

Recent work has shown that representational alignment between biological and artificial neural networks is asymmetric: model representations predict neural responses much better than neural responses predict model representations. This asymmetry raises the question of whether representational geometry contributes to bidirectional representational alignment. We hypothesized that steering representational geometry during training can systematically influence bidirectional alignment. To test this hypothesis, we developed a computational framework that integrates spectral regularization with bidirectional predictivity analyses. As an initial demonstration, we evaluated our framework using self-supervised contrastive vision models. Steering the spectral geometry of the learned representations substantially increased reverse predictivity with modest reductions in forward predictivity, yielding a 55% relative improvement in bidirectional predictivity. These improvements were accompanied by reduced effective dimensionality and a reorganization of the shared representational subspace, within which forward and reverse predictivity became approximately symmetric at intermediate spectral exponents. Overall, these findings demonstrate that representational geometry can be systematically steered to modulate bidirectional representational alignment between biological and artificial neural networks.

Samuel Kostousov, Abhinn Kaushik, Brokoslaw Laschowski

Department of Physics, University of Toronto, Toronto, ON, Canada · Department of Computer Science, University of Toronto, Toronto, ON, Canada · Department of Mechanical and Industrial Engineering, University of Toronto, Toronto, ON, Canada; and the KITE Research Institute, University Health Network, Toronto, ON, Canada

Neural Representation Geometry

Ask PDF

Abstract

Representation Identifiability

Explore similar work

Representation Alignment Rests on Linear Structure

Stimulus symmetries can confound representational similarity analyses

Bidirectional representational alignment between biological and artificial neural networks