TLDR: This research introduces a framework to systematically evaluate representational similarity metrics based on their ability to differentiate between various AI model families, including different architectures and training methods. By using measures like d-prime, silhouette coefficients, and ROC-AUC, the study found that metrics imposing stronger alignment constraints, such as Representational Similarity Analysis (RSA) and SoftMatch, are more effective at distinguishing model families than looser methods like linear predictivity. The findings provide guidance for selecting appropriate similarity metrics in AI and neuroscience research.
In the rapidly evolving fields of neuroscience and artificial intelligence, understanding how different models represent information is crucial. Researchers often use ‘representational similarity metrics’ to compare these models, but a significant challenge has been the lack of a systematic way to evaluate how well these metrics can actually tell different types of models apart. This new research, titled Measuring the Measures: Discriminative Capacity of Representational Similarity Metrics Across Model Families, addresses this gap by introducing a comprehensive framework to assess the discriminative power of these similarity measures.
The study, conducted by Jialin Wu, Shreya Saha, Yiqing Bo, and Meenakshi Khosla from UC San Diego, highlights a critical issue: with over a hundred comparison methods available, many produce inconclusive results. This methodological confusion can lead to misleading conclusions, such as the idea that architectural differences between CNNs and Transformers have negligible effects on brain alignment, which might simply be due to the analytical tools lacking the necessary sensitivity.
A Systematic Approach to Evaluation
The researchers set out to systematically benchmark the metrics themselves, rather than just using them to benchmark models. They analyzed 35 vision models, encompassing a wide range of architectures like Convolutional Neural Networks (CNNs), Vision Transformers, Swin Transformers, and ConvNeXt. These models also represented different training paradigms: supervised and self-supervised learning.
Four commonly used representational similarity metrics were put to the test: Representational Similarity Analysis (RSA), Soft Matching, Procrustes Alignment, and Linear Predictivity. These metrics vary in how flexibly they allow representations to be aligned. RSA, for instance, compares the relational structure of representations, while linear predictivity seeks any linear mapping to predict one representation from another, offering the most flexibility.
Measuring Separability
To quantify how well each metric could distinguish between model families, the team employed three complementary separability measures:
- d-prime: A measure from signal detection theory that quantifies the separation between within-family and between-family similarity distributions. Higher values indicate better separation.
- Silhouette Score: This score indicates how well models are clustered within their own families and separated from others. Scores near 1 mean good clustering, while negative values suggest misclassification.
- ROC-AUC: Treating family separability as a binary classification problem, ROC-AUC provides a robust global measure of discriminability, with values closer to 1 indicating perfect separation.
Key Findings: Constraints Lead to Clarity
The evaluation revealed clear differences in the discriminative capacity of the metrics. A crucial principle emerged: metrics that impose more stringent alignment constraints generally achieve superior discrimination. RSA, a non-fitting method that compares relational structure, demonstrated the strongest overall separability across all measures. Soft Matching, which finds an optimal probabilistic correspondence between units, followed closely, outperforming Procrustes Alignment and Linear Predictivity among the mapping-based approaches.
Specifically, RSA achieved the highest d-prime (3.79), silhouette score (0.51), and ROC-AUC (0.912). Soft Matching also showed strong performance, particularly in ROC-AUC (0.909). In contrast, Linear Predictivity, the most flexible metric, consistently showed the weakest discrimination across all measures.
This finding challenges the common assumption that looser metrics are better at capturing representational differences. Instead, the constraints imposed by metrics like RSA and SoftMatch appear to filter out incidental variations, allowing them to preserve the essential computational signatures that differentiate model families. Remarkably, even subtle distinctions, such as between supervised and unsupervised variants within the same architectural family, were reliably separated by RSA and SoftMatch.
Also Read:
- Rethinking Distance Measures in Representation Learning
- Evaluating AI: Bridging the Gap Between Benchmarks and Human Understanding
Guiding Future Research
The implications of this research are significant for both AI and neuroscience. It provides practical guidance for representation analysis, suggesting that researchers should select metrics based on their specific discrimination goals rather than simply defaulting to popular choices. By offering a principled framework to benchmark similarity metrics and clarify their trade-offs, this work paves the way for more interpretable and goal-aligned comparisons across diverse models and even between artificial and biological vision systems.
