Topological RSA (tRSA)

• tRSA is a framework that integrates geometric and topological transforms to compress and analyze representational dissimilarities in noisy, high-dimensional data.
• It employs methods such as Geo-Topological transforms, AGTDM, and pMDS to reveal dynamic structures and achieve high identification accuracies in diverse applications.
• The approach enhances noise robustness and interpretability in fields like neuroscience, DNN analysis, and single-cell studies while maintaining compatibility with classical RSA pipelines.

Topological Representational Similarity Analysis (tRSA) is a framework that systematically integrates geometric and topological properties in the characterization of neural or high-dimensional data representations. Originating as an extension to classical Representational Similarity Analysis (RSA), tRSA is centered on the transformation and compression of representational distances via nonlinear, monotonic mappings, thus enabling robust model comparison and interpretability even in the presence of noise or individual variability. Applications span neuroscience, artificial intelligence, and systems biology, providing a theoretically grounded approach to uncovering computational signatures, dynamical trajectories, and higher-order structures in complex datasets (Lin, 2024).

1. Mathematical Foundations and Geo-Topological Transforms

tRSA generalizes the classical RSA approach by introducing the Geo-Topological (GT) transform, which operates on the representational dissimilarity matrix (RDM), $D = (d_{ij})$, where $d_{ij}$ quantifies pairwise dissimilarity (e.g., correlation distance or cross-validated Mahalanobis distance) among representations $X_1, \dots, X_n$. For thresholds $0 \leq l < u \leq d_{\max}$, $d_{\max} = \max_{i,j} d_{ij}$, the GT transform is defined as

$$f_{l,u}(d_{ij}) = \begin{cases} 0, & d_{ij} \leq l \ \frac{d_{ij} - l}{u - l}, & l < d_{ij} < u \ 1, & d_{ij} \geq u \end{cases}$$

Each entry of $D$ is replaced by $f_{l,u}(d_{ij})$ to yield a Geo-Topological Matrix (GTM), $G^{(l,u)}_{ij}$. The family of GTMs, indexed by $(l,u)$, interpolates between full geometric structure ($l=0, u=d_{\max}$) and pure topological adjacency ($l \approx u$). This partial compression of very small and very large distances de-emphasizes noise and idiosyncratic features while highlighting robust neighborhood and intermediate-scale relationships.

Any GTM can substitute for the original RDM in RSA inference pipelines, supporting procedures based on Spearman’s $\rho$, Kendall’s $\tau_a$, cosine similarity, or bootstrap/permutation-based model comparison (Lin, 2024).

2. Adaptive Geo-Topological Dependence Measure (AGTDM)

tRSA’s GT transforms directly inform the Adaptive Geo-Topological Dependence Measure (AGTDM), which is designed to test statistical dependence between multivariate random vectors $X \in \mathcal{X}$ and $Y \in \mathcal{Y}$. AGTDM generalizes distance covariance and correlation by replacing entrywise distances with their GT-transformed counterparts. The metric is defined as

$$\mathcal{V}^{*2}(X, Y; l,u) = \mathbb{E}[f_{l,u}(\|X - X'\|) f_{l,u}(\|Y - Y'\|)] + \cdots - 2\mathbb{E}[f_{l,u}(\|X - X'\|) f_{l,u}(\|Y - Y''\|)]$$

and

$$dCor^*(X, Y; l, u) = \frac{\mathcal{V}^{*2}(X, Y; l, u)}{\sqrt{\mathcal{V}^{*2}(X, X; l, u) \mathcal{V}^{*2}(Y, Y; l, u)}}$$

The AGTDM statistic is the maximum $dCor^*(X, Y; l, u)$ over all feasible $(l, u)$. It is assessed via permutation testing by shuffling data to generate a null distribution, controlling type I error. AGTDM exhibits increased sensitivity and power for diverse dependency structures—including linear, polynomial, circular, checkerboard, and spiral patterns—compared to classical metrics (dCor, HSIC, MIC), particularly across varying noise levels and sample sizes (Lin, 2024).

3. Procrustes-aligned Multidimensional Scaling (pMDS) for Dynamics

To resolve time-evolving changes in neural or high-dimensional representational geometry, Procrustes-aligned Multidimensional Scaling (pMDS) is introduced. The method involves the following:

1. Construction of time-resolved RDMs $D^{(t)}$ at each time point (or sliding temporal window).
2. Independent application of classical MDS to each $D^{(t)}$ to obtain low-dimensional embeddings $Z^{(t)}$.
3. Temporal alignment of embeddings by Generalized Procrustes Analysis: minimizing $\|Z^{(t+1)}R + \mathbf{1}T - Z^{(t)}\|_F$ over orthonormal transformations $R$ (rotation, reflection) and translation $T$.
4. Visualization as a smooth trajectory, where the 1–2 axes encode geometry and the third axis or marker annotation encodes time (Lin, 2024).

In empirical studies, pMDS has elucidated dynamics in monkey inferotemporal (IT) cortex (showing divergence and convergence of object-category codes on subsecond timescales) and in DNN representations across computational layers or timesteps.

4. Temporal and Single-Cell Topological Analyses

tRSA’s scope is broadened with Temporal Topological Data Analysis (tTDA) and Single-cell Topological Simplicial Analysis (scTSA):

• Temporal Filtration: Data points $(x_i, t_i)$ are structured using both spatial and temporal thresholds $(\epsilon, \tau)$. Graph edges are formed when $\|x_i-x_j\|\leq\epsilon$ and $|t_i-t_j|\leq\tau$. The ensuing filtered simplicial complexes enable persistent homology calculations, producing temporal–topological invariants (e.g., Betti numbers, barcodes) with tunable time resolution. This approach captures the dynamic formation and extinction of clusters, cycles, and voids.
• scTSA: Applied to single-cell RNA-seq data, scTSA samples a subset of cells at each developmental time via a max-min scheme, computes correlation distances, and builds Vietoris–Rips complexes up to dimension $k$ (e.g., $k\leq7$). For each time point $t$, the number of $k$-simplices $SC_k(t)$ is normalized against a permutation-based null model to yield $NSC_k(t)$. In zebrafish embryogenesis, this method pinpointed gastrulation (5–6 hpf) as the most topologically complex developmental transition (Lin, 2024).

5. Comparison with Classical RSA

tRSA systematically contrasts with classical RSA, as summarized below:

Approach	Core Matrix	Sensitivity	Robustness
Classical RSA	RDM $D_{ij}$	Sensitive to all pairwise distances, including noisy/idiosyncratic ones	Sensitive to nuisance variation
tRSA	GTM $G^{(l,u)}$	Focuses on neighborhood graph at scale $(l,u)$, preserves computationally salient structure	Compresses nuisance variation

Empirical evaluations demonstrate that tRSA with intermediate $(l,u)$ achieves region-identification or layer-identification accuracies matching or exceeding classical RSA—76.5–77.2% in human fMRI region identification, and upwards of 95–100% in DNN layer identification (depending on noise level)—while being less sensitive to spurious distances (Lin, 2024).

6. Representative Case Studies

• Human fMRI: On data from Walther et al. (2016), tRSA achieved region-identification accuracies comparable to classical RSA (RIA ≈76.5% for RSA, ≈77.2% for tRSA at $(l ≈ 0.40, u ≈ 0.65)$). Both topology-sensitive and geometry-sensitive GTMs outperformed local-only descriptors.
• DNNs: In All-CNN-C networks trained on CIFAR-10, tRSA achieved high layer identification accuracy (LIA ≈97–100% at low noise, >90% at higher noise), matching or exceeding classical RSA performance. MDS and pMDS visualizations detailed the representational evolution across model depth and noise parameters.
• Monkey IT Dynamics: pMDS of sliding-window RDMs highlighted divergence of object representations at 100 ms post-stimulus, peaking at 150 ms and then convergence after stimulus offset, congruent with neurophysiological findings.
• Zebrafish scRNA-seq: scTSA revealed a sharp increase in higher-order simplices at gastrulation, accentuating the topological complexity associated with critical developmental transitions. Mapper applied with temporal filtration resolved bifurcation in developmental trajectories (Lin, 2024).

7. Interpretive Advantages, Limitations, and Outlook

tRSA and ancillary methods deliver several key advantages:

• Robustness: Compression of extreme distances suppresses nuisance variation while retaining computational signals.
• Flexibility: The parameter space $(l,u)$ offers continuous tuning between geometric and topological sensitivity.
• Unified Inference: Existing RSA pipelines are compatible, and methods generalize to geodesic distances and beyond.
• Extension to Dynamics and Nonlinear Dependencies: pMDS and tTDA accommodate temporal structure; AGTDM provides an adaptive, powerful multivariate dependence test.
• Applicability: Framework encompasses a broad range of high-dimensional domains.

Limitations include the need for 2D parameter optimization over $(l,u)$ (necessitating careful validation), computational expense for large $m$ or $N$, and conceptual challenges in the interpretation of high-order and multi-parameter topological features.

Future directions entail applications to disease states ("neural dysmanifolds"), unification with information-theoretic metrics, detailed joint analysis across $(l,u)$ as a two-dimensional persistence module, integration with kernel/graphical independence testing, and leveraging tRSA principles to direct DNN design towards models with biologically plausible topological biases (Lin, 2024).