When we train two different neural networks on the same task, do they learn the same internal representations? To answer this, researchers traditionally rely on statistical similarity metrics like Centered Kernel Alignment (CKA) or Canonical Correlation Analysis (CCA). These metrics look at the internal activations of two networks and output a similarity score based on their geometric alignment.
However, a growing body of research reveals a fundamental flaw in this approach: mathematical similarity does not equal functional equivalence. Two networks can behave identically while appearing completely different to a similarity metric. Conversely, a metric can rate two representations as highly similar even when they produce entirely different behaviors.
To build reliable models and truly understand deep learning, we must shift from purely statistical metrics toward the functional grounding of representational similarity—evaluating internal representations based on how they actually drive a network's behavior.
The Geometry-Function Disconnect
Why do standard statistical metrics fail to capture true representational alignment? The core issue is that metrics like CKA evaluate the global geometry of activation spaces rather than how that geometry is utilized by downstream layers.
Consider a network whose internal representations are altered by a simple geometric transformation, such as translating a subset of data points or applying an invertible linear transformation. To a downstream classifier, the underlying structure and linear separability of the data remain completely unchanged; the network's final accuracy is unaffected. Yet, to a statistical metric, the representation looks entirely different.
[Input Data] ---> [Layer Activations] ---> [Downstream Classifier] ---> [Same Accuracy]
|
(Affine Transformation)
v
[Transformed Activations] ---> [CKA Score Drops Drastically]
This fragility was exposed by MohammadReza Davari et al., who demonstrated that CKA is highly sensitive to simple data translations. By shifting subsets of data points, the authors could drastically manipulate CKA scores without altering the network's functional performance or the linear separability of its features. In a companion study, MohammadReza Davari et al. showed that this sensitivity extends to broader affine transformations. Because CKA can be easily deceived by transformations that are behaviorally inconsequential, it frequently yields misleading conclusions, reporting low similarity for functionally identical layers.
Proving the Dissociation
This gap between geometry and behavior is not just an artifact of fragile metrics; it is a fundamental property of neural network optimization. Internal representations can analytically decouple from the network's input-output function.
Using deep linear networks, Lukas Braun et al. mathematically proved this dissociation. The authors showed that many distinct weight configurations can achieve the exact same input-output mapping while hosting wildly different internal codes. Some of these codes reflect the intrinsic structure of the task, while others appear completely arbitrary. As a result, high representational similarity can occur between networks performing different tasks, and functional equivalence can easily exist without representational similarity.
/---> Internal Code A (Task-Structured) ---\
[Input Function] ---> [Same Output Behavior]
\---> Internal Code B (Arbitrary Noise) ---/
We can observe this dissociation empirically through a technique called "model stitching," where the first half of one network is directly connected to the second half of another via a simple linear layer. If two networks have learned functionally equivalent features, the stitched hybrid model should perform well.
When Adrián Csiszárik et al. performed model stitching across networks trained from different random initializations, they discovered that the models were remarkably compatible. Crucially, the hybrid models maintained high task accuracy even when standard statistical metrics like CKA explicitly minimized or failed to predict their compatibility. The internal features were functionally interchangeable, even though they were mathematically dissimilar.
From a philosophical standpoint, this challenges recent assertions that AI models are naturally converging on a single, unified representation of reality. As Gilad Landau et al. argue, relying on kernel alignment or structural covariance is insufficient for establishing true representational content. Just because two activation spaces share a statistical structure does not mean they represent the world in the same way; we require a deeper functional analysis.
Grounding Metrics in Behavioral Reality
If purely statistical measures are unreliable, how can we validate whether two representations are truly similar? The solution is to ground similarity metrics in functional correspondence—evaluating the metric itself by how well its scores correlate with actual changes in model behavior.
[Metric Distance Increases] <===(Strong Correlation)===> [Model Accuracy Drops]
Frances Ding et al. introduced a rigorous framework to test this relationship, evaluating metrics based on whether an increase in representational "distance" reliably predicts a drop in model accuracy. Their findings revealed that while popular metrics like CKA and CCA possess significant blind spots, the classical Orthogonal Procrustes distance stands out as a robust performer. Because Procrustes explicitly aligns representations via an orthogonal rotation before measuring distance, it bypasses many of the inconsequential geometric shifts that deceive other metrics.
Similarly, Yiqing Bo et al. evaluated representational similarity measures through the lens of behavioral reality. The authors confirmed that global geometric metrics like Procrustes distance serve as superior indicators of functional similarity. Conversely, they found that common neuroscience benchmarks, such as linear predictivity, often fail to capture meaningful behavioral differences, reinforcing the danger of relying on metrics detached from downstream performance.
The Path Forward
The core challenge for the field is to design next-generation metrics that can separate structural noise from functionally relevant representational shifts. As Chaitanya Kapoor et al. note, representational alignment often occurs remarkably early in training, driven heavily by architectural biases and data statistics rather than final task mastery. Current measures struggle to differentiate between these early, passive structural alignments and the active, functional representations that dictate a model's final output behavior.
By moving away from purely geometric comparisons and demanding that similarity metrics be grounded in functional behavior, researchers can develop a more accurate, deceptive-resistant understanding of what neural networks actually learn.
Key Foundational Work
On the Inadequacy of CKA as a Measure of Similarity in Deep Learning (2022) — contributed 10 concept elements to this direction.
Deceiving the CKA Similarity Measure in Deep Learning (2022) — contributed 4 concept elements to this direction.
Similarity and Matching of Neural Network Representations (2021) — contributed 3 concept elements to this direction.