The Incommensurability Problem
How do you compare what happens inside a deep neural network layer with 4,096 dimensions to what happens inside a human visual cortex measured via fMRI voxels? Or how do you determine if two different language models, trained on different data with different architectures, have learned the same underlying concepts?
Direct point-to-point mapping is usually impossible. The coordinate systems, dimensionalities, and fundamental architectures of these systems are completely different. As noted by SUBBA REDDY OOTA in Deep Neural Networks and Brain Alignment: Brain Encoding and Decoding (Survey) and Domenic Bersch et al. in Net2Brain: a toolbox to compare artificial vision models with human brain responses, evaluating these systems requires a method that abstracts away from specific unit-to-neuron or coordinate-based mappings.
The Core Intuition: Relational Geometry
The breakthrough idea is to shift our focus from absolute coordinates to relational geometry. Instead of asking where a specific concept (like a "dog") is located in a given vector space, we ask how a system differentiates a dog from a cat, a car, or a tree.
By looking at the relative distances between representations of a shared set of stimuli, we can build a universal bridge between entirely different systems. This framework, widely known as Representational Similarity Analysis (RSA), allows us to compare information structures directly.
As explained by Adrien Doerig et al. in The neuroconnectionist research programme, focusing on the relative distances between representations allows researchers to assess functional similarity without requiring a direct one-to-one mapping between individual artificial units and biological neurons.
How it Works: The Representational Dissimilarity Matrix
To understand why this works, consider the step-by-step process of constructing a Representational Dissimilarity Matrix (RDM), as detailed by Johannes Singer et al. in From photos to sketches - how humans and deep neural networks process objects across different levels of visual abstraction and Jason Gors et al. in The Representation of Biological Classes in the Human Brain:
- Universal Anchors: We select a shared set of $N$ input stimuli (such as a collection of sounds, images, or sentences) and present them to both systems.
- Pairwise Distances: For System A (e.g., an fMRI scan of the auditory cortex), we record the internal activation pattern for every stimulus. We then calculate the mathematical distance (or dissimilarity) between every possible pair of stimuli. This forms an $N \times N$ matrix where each entry $(i, j)$ represents how distinctly the system treats stimulus $i$ versus stimulus $j$.
- Matrix Comparison: We repeat this process for System B (e.g., a layer of a deep neural network).
Now, we have two matrices of identical dimensions ($N \times N$), regardless of the original dimensionality or nature of System A and System B. We can simply correlate these two matrices to find out how well their internal organizational principles align.
[Stimuli Set] ---> Present to System A (e.g., Brain) --------> Compute Pairwise Distances ---> [ RDM A (NxN) ]
---> Present to System B (e.g., AI Model) -----> Compute Pairwise Distances ---> [ RDM B (NxN) ]
|
Correlate RDMs
|
v
Representational Alignment
Early implementations of this concept include Haxby's correlation technique, which Leila Reddy utilized in Category selectivity in the ventral visual pathway confers robustness to clutter and diverted attention to measure the similarity between fMRI response patterns evoked by different object categories.
Why the Geometry Matters: Topology Over Coordinates
Why does this non-parametric approach work so well? Because it preserves the underlying structural topology. As Ilia Sucholutsky et al. show in Alignment with human representations supports robust few-shot learning, measuring representational alignment via relative distances captures the structural topology of the representation rather than absolute coordinate values.
This allows us to uncover hidden structures in ungrounded models. For instance, Mostafa Abdou et al. demonstrated in Can Language Models Encode Perceptual Structure Without Grounding? A Case Study in Color that text-only language models implicitly reconstruct the 3D topological relationships of human color perception.
Similarly, we can compare a model's internal representations against predefined "hypothesis matrices" that represent the logical structure of a task. Safoora Yousefi et al. applied this in In-Context Learning in Large Language Models: A Neuroscience-inspired Analysis of Representations to quantify how closely an LLM's latent space aligns with specific conceptual or relational frameworks during in-context learning.
Applications Across Domains
Mapping Brains to AI
Umut Güçlü et al. used RSA in Brains on Beats to reveal a hierarchical gradient in the auditory cortex by comparing brain scans to music-tagging DNN layers. Colin Conwell et al. scaled this up in Neural Regression, Representational Similarity, Model Zoology & Neural Taskonomy at Scale in Rodent Visual Cortex to evaluate over 110 DNN architectures against thousands of rodent neurons, showing that trained models successfully recapitulate biological hierarchies.
Tracking Model Evolution and Drift
RSA can track how a model's internal space changes over time or across training paradigms. Amil Merchant et al. used it in What Happens To BERT Embeddings During Fine-tuning? to measure geometric alignment before and after fine-tuning. In multi-agent communication, Diane Bouchacourt et al. used RSA in How agents see things: On visual representations in an emergent language game to quantify how much an agent's internal conceptual hierarchy drifts from the original input structure when learning emergent languages.
Evaluating Cognitive Plausibility
By correlating model embeddings with human behavioral data or similarity ratings, benchmarks can determine if a network organizes concepts in a manner consistent with human perception. This strategy is central to the developmental evaluation benchmark proposed by Alvin Wei Ming Tan et al. in DevBench: A multimodal developmental benchmark for language learning and the auditory evaluation performed by Giancarlo Valente et al. in Semantically-Informed Deep Neural Networks For Sound Recognition.
Alternative and Complementary Approaches
While RSA relies on pairwise distance matrices, other tools offer alternative lenses for comparing representations:
- Canonical Correlation Analysis (CCA): Used by Yuanchao Li et al. in Exploration of A Self-Supervised Speech Model: A Study on Emotional Corpora, CCA quantifies the statistical similarity between high-dimensional hidden states and handcrafted, domain-specific features.
- Inference-Time Decomposition of Activations (ITDA): At massive scales where computing full distance matrices is expensive, Patrick Leask et al. introduced ITDA in Inference-Time Decomposition of Activations (ITDA): A Scalable Approach to Interpreting Large Language Models. This technique compares internal representations across different models by measuring the overlap of selected dictionary atoms when models are exposed to the same input prompts.