When we train two different neural networks to perform the exact same task—such as classifying images or predicting the next word in a sentence—do they learn the same internal representations?
To answer this, researchers traditionally use representational similarity metrics like Canonical Correlation Analysis (CCA) or Representational Similarity Matrices (RSMs). These tools look at the internal activations of a network and try to determine how aligned they are with another network's activations.
However, a growing body of research reveals a fundamental flaw in this approach: standard similarity metrics are highly sensitive to the physical "geometry" or structural container of a network, often failing to capture the actual underlying computation. To build truly robust comparison tools, we must move past raw spatial geometry and design metrics that are invariant to the mathematical symmetries inherent in neural architectures and data.
The Illusion of Geometric Dissimilarity
The core problem with traditional representational metrics is architectural bias. They frequently mistake structural properties—like activation functions, layer widths, or random initializations—for the actual computational logic of the model.
In a striking demonstration of this phenomenon, Maheswaranathan et al. (2019) trained thousands of recurrent neural networks (RNNs) across various architectures. They discovered that standard metrics like CCA are so sensitive to architectural signatures that they often judge an untrained model to be more similar to a trained model of the same architecture than two trained models of different architectures performing the exact same task. The metric measures the container, not the content.
This limitation was systematically evaluated by Bo et al. (2024) through the lens of functional correspondence—how well a similarity metric reflects actual behavioral alignment. They found that many popular benchmarks in neuroscience and machine learning fail to capture true functional similarity because they are confounded by these structural constraints.
Why Geometry Drifts: Symmetries and Gauge Redundancy
Why do functionally identical networks look so different geometrically? The answer lies in the mathematical symmetries present within neural networks and the data they process.
Consider a standard ReLU network. You can multiply the weights entering a hidden neuron by a factor of 10, and multiply the weights leaving it by 1/10. Because of the positive scaling property of ReLUs, the network's output remains completely unchanged. Similarly, you can permute the ordering of neurons within a layer without altering the network's function. As shown by Choi et al. (2021), these scaling and permutation symmetries mean that identical model outputs can be produced by internal representations with vastly different statistical signatures, such as varying levels of sparsity or correlation.
Symmetries in the input data compound this issue, creating what is known as gauge redundancy. As Pashakhanloo et al. (2025) demonstrate, when a stimulus space contains inherent symmetries, a network can tile that space using multiple, functionally equivalent neural codes. However, because traditional RSMs track the specific spatial arrangement of neural features rather than their functional coverage, these equivalent representations yield entirely different similarity matrices. The metric "drifts" even though the underlying computation is stable.
[Network A: Features arranged via Strategy 1] ──┐
├─► Identical Functional Output
[Network B: Features arranged via Strategy 2] ──┘
│
▼ (Traditional RSM / CCA)
"Highly Dissimilar" (Confounded by Spatial Arrangement)
Factoring Out Symmetries with Intertwiner Groups
To overcome these limitations, recent work aims to bake symmetry invariance directly into the comparison metrics. If a metric knows which transformations do not alter a network's function, it can ignore those transformations during comparison.
Godfrey et al. (2022) formalized this approach by introducing intertwiner groups—mathematical symmetry groups derived directly from a model's specific activation functions (such as permutations and positive scaling).
Instead of comparing raw activation vectors, symmetry-invariant metrics evaluate representations only after factoring out the transformations allowed by these intertwiner groups. When applying these group-aware measures, the authors revealed a hidden consistency: models that appeared distinct under traditional metrics actually possessed highly stable, shared representational structures across identical layers and residual blocks.
From Spatial Coordinates to Dynamical Flow
The challenge of representation comparison deepens when moving from static feedforward networks to recurrent networks that compute over time. In recurrent circuits, looking at the static geometry of activations is fundamentally insufficient because the computation is embedded in the system's time-evolving trajectories.
To solve this, Ostrow et al. (2023) proposed Dynamical Similarity Analysis (DSA). Instead of asking where representations sit in a high-dimensional space, DSA asks: What is the underlying vector field driving the state trajectories?
Using Koopman operator theory, DSA extracts the mathematical "flow" and temporal structure of a neural circuit. By comparing the dynamical operators rather than the physical coordinate embeddings, DSA can identify when two networks are utilizing the exact same computational strategy, even if they possess entirely different architectures, coordinate systems, or noise profiles.
Conclusion
The evolution of representational similarity analysis highlights a shift from shallow geometric alignment to deep functional and topological correspondence. By designing metrics that are invariant to architectural symmetries, data gauge redundancies, and coordinate embeddings, machine learning researchers can finally look past the superficial "quirks" of a network's structure to see the universal computational logic operating beneath.
Key Foundational Work
Universality and individuality in neural dynamics across large populations of recurrent networks. (2019) — contributed 3 concept elements to this direction.
Statistical Characteristics of Deep Representations: An Empirical Investigation (2021) — contributed 1 concept elements to this direction.