Topographical Deep Neural Networks (TDANNs)

• Topographical Deep Neural Networks (TDANNs) are deep learning architectures that integrate spatial and topological principles into network designs inspired by cortical organization.
• They leverage methods like topological data analysis, self-organizing maps, and spatial similarity losses to extract and regularize representational geometry.
• TDANNs improve model interpretability and brain-model alignment through explicit constraints and visualization tools that map activations onto structured grids.

Topographical Deep Neural Networks (TDANNs) constitute a diverse class of architectures, analysis methodologies, and tools that explicitly encode, exploit, or interpret the spatial, topological, or manifold-structured organization of representations within deep learning models. TDANNs are motivated both by biological observations of cortical topography and by the empirical and theoretical benefits of embedding topological ideas into network design, training, and analysis. Recent advances span supervised learning, generative modeling, neuroscience-inspired modeling, input space mapping, and principled brain-model comparison, uniting approaches from self-organizing map-based representations, topological data analysis (TDA), biologically inspired wiring, spiking dynamics, and explicit topology-aware objectives and constraints.

1. Foundations and Architectures

TDANNs arise from the integration of topographical and topological principles into neural network internal structure. Early proposals introduced multilayered restricted radial basis function networks (M-rRBFs) in which each hidden layer is a two-dimensional map with radial basis function activations and a neighborhood function enforcing locality. This advances traditional self-organizing maps (SOMs) by injecting top-down, label-driven error feedback into map formation, enabling class-specific homeomorphic clustering and "context-relevant" representations (Trappenberg et al., 2014).

More recent work replaces engineered weight-sharing (as in CNNs) with parameterizations over manifolds (e.g., circle, Klein bottle) or arbitrary graphs representing biological connectomes. For example, deep connectomics networks redraw the wiring of layers according to the DAGs induced by real neuronal networks, resulting in small-world properties, modularity, and improved generalization (Roberts et al., 2019). All-topographic neural networks (All-TNNs) forgo convolutional weight-sharing entirely, arranging units into explicit retinotopically organized sheets and enforcing spatial similarity losses between neighboring units to drive local smoothness and the emergence of cortical-like topographies (Lu et al., 2023).

Spiking implementations, such as TDSNNs, explicitly assign neurons to a virtual cortical sheet and apply losses correlating spatial proximity with similarity in firing rates and spike synchrony, achieving neural layouts featuring orientation pinwheels (V1) and category islands (IT), with high brain-likeness and negligible performance reductions compared to non-topographic models (Zhou et al., 6 Aug 2025).

2. Topological Data Analysis and Representational Geometry

A core methodological thread in TDANNs is the application of topological data analysis to neural activations and learned weights. Persistent homology, Mapper, and related TDA tools are used to quantify the "shape" of internal representations, revealing features such as clusters, loops, and cavities at different network depths.

For a given point cloud of activations $X^l$ at layer $l$, TDA constructions such as the Vietoris–Rips complex,

$$VR_{\varepsilon}(X^l) := \{[x_0^l,\dots,x_m^l] \mid d(x_i^l, x_j^l) \leq 2\varepsilon,\, \forall\, x_i^l, x_j^l \in X^l \},$$

support computing Betti numbers $\beta_k$ (topological feature counts) and persistent barcodes capturing the birth and death of topological features as distance thresholds grow. Mapper analyses extract coarse-grained graphs summarizing data geometry based on lens functions and local clustering (Goldfarb, 2018, Carlsson et al., 2018, Wheeler et al., 2021, Świder, 8 Jul 2024).

These analyses show:

• In deep classifiers, topological complexity generally increases with training: persistent features in activations become more pronounced, indicating accentuation of robust geometric invariants required for discrimination.
• Architectural differences (e.g., VGG19, ResNet, ViT) manifest as distinct topological evolution. For instance, ViT layers often display higher early outlier counts and more symmetric persistent feature distributions than CNNs (Świder, 8 Jul 2024).
• Pre-training versus fine-tuning: pre-trained and fine-tuned models exhibit similar topology in early layers but diverge in later layers, consistent with general feature extraction followed by task-specific specialization.

These properties can serve as diagnostic signals for representation analysis, model selection, and architectural regularizer development.

3. Learning Principles: Locality, Supervision, and Topographic Constraints

TDANNs unify unsupervised local feature aggregation and supervised task optimization. In M-rRBFs, each hidden map layer applies a combination of unsupervised competitive learning (as in SOMs) and error-modulated updates:

• Neuron activations,

$O_k^M(t) = \exp(-I_k^M(t))\,\sigma(k_*^M, k, t),$

where $$I_k^M(t)=\frac{1}{2}\|\mathcal{W}_k^M(t)-O^{M-1}(t)\|^2$$, and $\sigma$ is a neighborhood function.

• The reference vectors $\mathcal{W}_k^M$ are updated by bottom-up input similarity and top-down supervised error signals. The recursive error,

$$\delta_b^{(N-L)}(t) = (-1)^l \left(\sum_\alpha \Delta W_{\alpha b}^{(N-L+1)}(t)\right) e^{-I_b^{(N-L)}(t)},$$

propagates the supervised context into deep layers, refining the topographical map organization to favor contextually relevant clusters (Trappenberg et al., 2014).

Constraint terms proliferate in later work:

• Spatial similarity loss in cortical sheet architectures:

$$\mathcal{L}_\text{spat} = \frac{1}{N} \sum_{i,j}[\mathrm{cos\_dist}(w_{i,j},w_{i+1,j})+\mathrm{cos\_dist}(w_{i,j},w_{i,j+1})],$$

penalizing abrupt feature variation between spatial neighbors (Lu et al., 2023).

• Spatio-temporal constraints in SNNs:

$$\mathcal{L} = \mathcal{L}_\text{task} + \frac{1}{M}\sum_{k=1}^K\sum_{m=1}^M[\alpha \mathcal{L}_L(k,m) + \beta \mathcal{L}_S(k,m)],$$

$\mathcal{L}_L$ promoting long-term firing rate correlation among nearby units, $\mathcal{L}_S$ promoting short-term spike synchrony (Zhou et al., 6 Aug 2025).

• Wiring cost constraints in TDANNs that penalize large Euclidean distances between similarly tuned neurons:

$$L_\text{total} = L_\text{classification} + \lambda \sum_{i,j} s(R_i, R_j)\|p_i - p_j\|^2,$$

where $s(R_i, R_j)$ is a similarity metric between response profiles (Lee et al., 2019).

These objective components guide the emergence of spatial-functional organization, enable specialization (e.g., face patches), and can be tuned to avoid performance penalties commonly observed when imposing topographic constraints.

4. Practical Applications and Empirical Findings

TDANNs have demonstrated empirical advantages across tasks:

• Semantic mapping: TopoNets, using SPN templates distributed over dynamic topological graphs, enable tractable real-time semantic inference, outperforming MRF and local SPN baselines in both explored and unexplored space classification accuracy in robotic mapping (Zheng et al., 2018).
• Visual cortex modeling: Imposing minimum wiring cost and spatial constraints yields networks that explicitly match the spatial clustering and fall-off properties of category selectivity (e.g., face patches) in primate inferior temporal cortex (Lee et al., 2019, Zhou et al., 6 Aug 2025, Lu et al., 2023). TDSNNs closely reproduce the hierarchy from V1 orientation maps and "pinwheel" structures to IT-like category patches, with high BrainScore alignment and negligible or improved classification accuracy relative to unconstrained baselines (Zhou et al., 6 Aug 2025).
• Diffusion models: TopoDiffusionNet incorporates persistent homology-based losses to directly enforce desired Betti number (topological structure count) constraints in generative denoising, enabling accurate mask generation with prescribed numbers of objects or holes. Loss terms $\mathcal{L}_{\text{preserve}}$ and $\mathcal{L}_{\text{denoise}}$ sharpen or suppress individual persistent structures in the intermediate output, dramatically improving topological precision compared to conditioning-based or baseline diffusion approaches (Gupta et al., 22 Oct 2024).
• Input space mapping and testing: TopoMap constructs an explicit topographical map of the input feature space using black-box embedding + clustering pipelines, with DNN-driven cluster discriminability evaluation. In mutation analysis, regions selected from these maps outperform random selection by 35% on average in killable mutant detection, demonstrating improved prioritization for model robustness testing (Vita et al., 3 Sep 2025).

5. Visualization and Interpretability

TDANNs facilitate advanced interpretability tools by providing natural spatial substrates over which activation or error patterns can be visualized and diagnosed.

• Topographic activation maps layout neurons in two-dimensional grids, aligning spatial proximity in the layout with similarity in activation profiles across data groups or classes. Methods such as SOMs, co-activation graphs, or particle swarm optimization produce layouts that aid in error diagnosis, bias exposure, and visualization of training dynamics (Krug et al., 2022).
• Persistent landscapes and Mapper graphs summarize high-dimensional neural activations, revealing persistence of structural features as learning proceeds and enabling statistical kernel methods for further meta-analyses (e.g., hypothesis testing on network evolution across layers) (Wheeler et al., 2021, Goldfarb, 2018).

This suite of post-hoc and online visualization methods connects the representational geometry of TDANNs to the explainability demands of modern ML systems, bridging conceptual and practical interpretability.

6. Model–Brain Alignment and Mechanistic Comparison

TDANNs are increasingly positioned as mechanistic models of neural circuitry, especially in vision. Inter-Animal Transform Class (IATC) methodology enables principled, bidirectional mappings between model activations and brain responses by restricting transforms to those needed to align real neuronal populations. The key "zippering transform" explicitly inverts model or biological nonlinearities before performing a fitted mapping and then re-applies the nonlinearity:

$$\text{Pre-activation:}\ X_\text{pre} = \sigma^{-1}(X_\text{post}) \ \text{Linear map:}\ Y_\text{pre} = W \cdot X_\text{pre} \ \text{Post-activation:}\ Y_\text{post} = \sigma(Y_\text{pre})$$

This approach achieves both high predictivity and specificity—preserving separations between distinct brain areas. Applying IATC-guided transforms reveals that topographic models with intermediate spatial loss strengths most faithfully reproduce brain-like response hierarchies and offer improved selectivity compared to non-topographic counterparts (Thobani et al., 2 Oct 2025).

7. Theoretical Perspectives: Topological Classification and Architectural Implications

A unifying formalism posits classification as a topological separation problem: given a topologically structured input space $X$ and a labeling map $g: X \to \mathcal{Y}$, a classifier $f: X \to Y$ is well-posed only if

$$f(g^{-1}(y)) \subset U_y \text{ for disjoint open sets } U_y \subset Y$$

for all $y \in \mathcal{Y}$. Classical topological lemmas ensure a continuous $f$ exists only if the class supports are topologically separable. TDANN architectures may thus be designed (e.g., in early layers) to "unfold" nontrivial data topology and facilitate such separability, informing choices about depth, width, and permissible non-linearities (Hajij et al., 2021). Persistent homology and Betti number profiles provide a quantitative means of verifying when and where these separations are achieved.

---

TDANNs have thus evolved from early context-sensitive topographical maps to a hyper-general class of architectures and analyses infused with topological and geometric reasoning. Their key feature is the operationalization of spatial or topological induction—whether via direct map constraints, biologically inspired wiring, topology-aware losses, or the interpretive application of TDA. They serve as both models of brain function and as practical, empirically robust learning systems, with increasing attention paid to their unique advantages in robustness, generalization, interpretability, and neuroscience alignment.