Topographic Convolutional Neural Networks

• Topographic Convolutional Neural Networks are models that incorporate explicit spatial organization into CNN layers, arranging channels or inputs on geometric grids to mimic neurobiological topography.
• They utilize techniques such as channel lattices, topographic regularization, and spatial mapping of irregular data (e.g., EEG or terrain) to improve feature robustness and connectivity.
• These models enhance performance under challenging conditions like channel pruning and input degradation, and they are adapted for tasks ranging from biological map formation to physical topography reconstruction.

Searching arXiv for recent and foundational papers on topographic convolutional neural networks and closely related formulations. arXiv search query: "topographic convolutional neural networks topography CNN topographic representation module introducing topography in convolutional neural networks" Topographic Convolutional Neural Networks designate a heterogeneous family of convolutional models in which “topographic” refers to an explicit spatial organization of representations, signals, or targets. In one line of work, convolutional channels or units are placed on a geometric lattice and regularized so that nearby units develop similar responses, approximating neurobiological topography. In another, irregular measurements such as EEG channels are mapped to a spatial topographic grid before convolution. In a third, CNNs are used to predict or reconstruct physical topography, including terrain, coastal profiles, fracture surfaces, and road-network structure. Taken together, these studies indicate that the term does not denote a single canonical architecture, but rather a set of inductive biases and problem formulations in which spatial organization is treated as a first-class design principle (Poli et al., 2022).

1. Terminological scope and problem classes

The literature uses the term across several technically distinct settings. One setting introduces a channel-level geometric organization inside standard CNNs: output channels of a convolutional layer are assigned positions in a 1D, 2D, or 3D embedding, and training encourages nearby channels to exhibit similar activations. Another setting treats topography as an input representation problem, as in EEG decoding, where electrode measurements are mapped onto a 2D scalp layout so that convolution kernels operate on actual spatial neighborhoods rather than arbitrary channel order. A third setting uses CNNs to estimate topographic structure in the external world, such as beach elevation profiles, local terrain geometry for locomotion, or digital elevation models. A fourth setting applies CNNs to topology-sensitive reconstruction problems, such as road and vessel tracing, where the central issue is connectivity rather than scalar height.

Sense of “topographic”	Representative formulation	Representative papers
Channel or unit organization	Lattice over channels or cortical sheet with similarity regularization	(Poli et al., 2022, Truong et al., 31 Jul 2025, Lee et al., 2019, Lu et al., 2023)
Spatialized input representation	Mapping irregular sensors to a 2D grid before CNN processing	(Liang et al., 2022)
Physical surface prediction	CNNs or hybrid conv models infer terrain, fracture, or coastal topography	(Tsopanidis et al., 2019, Hoeller et al., 2022, Munian et al., 11 Jan 2025, Costante et al., 2018)
Cartographic rendering or recognition	CNNs render or classify map content tied to terrain depiction	(Jenny et al., 2020, Zhou et al., 2018)
Connectivity-aware tracing	Patch CNNs predict local connectivity and reconstruct global networks	(Ventura et al., 2018)

This breadth has two consequences. First, comparisons across papers must be made at the level of inductive bias rather than task label alone. Second, “topographic CNN” in contemporary usage can describe either an architectural prior inside the network or a CNN applied to data whose primary semantics are topographic.

2. Channel lattices and topographic regularization

A central architectural formulation appears in “Introducing topography in convolutional neural networks” (Poli et al., 2022). There, standard CNN channels are no longer treated as unordered slots. For each convolutional layer $l$, output channel $i$ is assigned a position $p_i^{(l)}$ in a topographic embedding space, with supported organizations including a 2D regular grid, 2D nested grid, 2D circle, 3D regular grid, 3D nested grid, and 3D sphere. Similarity is computed in activation space, using cosine similarity between flattened pre-activation outputs:

$$S^{(l)}_{i,j}(y^{(l)}) \triangleq \frac{\sum_{h=1}^{H} \sum_{w=1}^{W} y^{(l)}_{i,h,w}\, y^{(l)}_{j,h,w}} {\left\| \tilde{y}^{(l)}_i \right\|_2 \left\| \tilde{y}^{(l)}_j \right\|_2},$$

with target similarity profile

$$\tau^{(l)}_{i,j} \triangleq \frac{1}{d^{(l)}(i,j)+1},$$

where $d^{(l)}(i,j)=\|p_i^{(l)}-p_j^{(l)}\|_2$. The layer-wise loss is

$$L^{(l)} \triangleq \frac{2}{\text{out}(\text{out}-1)} \sum_{1 \le i < j \le \text{out}} \left(S^{(l)}_{i,j}-\tau^{(l)}_{i,j}\right)^2,$$

and the total objective is $$\mathcal{L}=\mathcal{L}_{\text{task}}+\lambda \cdot \left(\frac{1}{L}\sum_{l=1}^{L}L^{(l)}\right)$$.

The formulation is architecture-agnostic: it was applied to ResNet-18, VGG-16, and DenseNet-121, with topography placed on all convolutional layers in VGG-16 and on block-end convolutions in ResNet-18 and DenseNet-121. Across CIFAR10, CIFAR100, SpeechCommands, and BirdDCASE, topographic variants matched baseline performance, with modest gains in some configurations. Its clearest empirical advantage appeared under magnitude-based structured pruning: on ResNet-18 for CIFAR10, without pruning the topographic model reached 91.9% versus 92.45% for the baseline, but under 62% channel pruning per topographic layer it retained 62.0% test accuracy versus 40.6% for the baseline. The paper attributes this to smoother, redundant representations over nearby channels.

A later comparison between local Weight Similarity (WS) and Activation Similarity (AS) constraints shifts the regularizer from activation space to incoming weights (Truong et al., 31 Jul 2025). In that work, a topographic fully connected layer $fc1$ is arranged as an $11 \times 11$ grid, and the WS loss averages Euclidean distances between neighboring incoming weight vectors:

$i$0

The AS alternative averages correlation distance between neighboring activation vectors. On MNIST and CIFAR-10, WS produced improved robustness to weight corruption and input degradation, higher activation variance, and stronger functional localization than AS or non-topographic controls. The paper reports that WS yielded shorter distances between similarly responding units, positive Moran’s $i$1 increasing with $i$2, and smaller accuracy drops under weight perturbation; by contrast, AS produced bimodal pairwise activation correlations and negative Moran’s $i$3, described as “striped” patterns. This distinction is important: topography can be enforced either by constraining responses directly or by smoothing the weight manifold that generates those responses.

3. Cortical-sheet models, map formation, and lateral interactions

A biologically grounded branch of the literature imposes topography over cortical sheets rather than over abstract channel lattices. “Topographic Deep Artificial Neural Networks (TDANNs) predict face selectivity topography in primate inferior temporal (IT) cortex” places units in AlexNet fc6 on a 2D tissue map and regularizes pairwise response correlations as a function of cortical distance, using a macaque-derived correlation–distance curve as target (Lee et al., 2019). The objective takes the form

$i$4

with $i$5 penalizing deviations between model pairwise correlation and the empirical curve $i$6. The resulting models produce clustered face-selective maps whose purity fall-off with cortical distance resembles macaque middle face patches.

“End-to-end topographic networks as models of cortical map formation and human visual behaviour: moving beyond convolutions” removes weight sharing entirely and replaces convolutional layers with locally connected layers organized on a 2D cortical sheet (Lu et al., 2023). Each layer is regularized by a neighborhood smoothness term over kernel vectors:

$i$7

The architecture comprises six locally connected layers, with layers 1, 3, and 5 followed by $i$8 pooling, layer normalization, ReLU, $i$9 regularization $p_i^{(l)}$0, dropout 0.2, and training on ecoset for 600 epochs with Adam at learning rate 0.001 and $p_i^{(l)}$1. Several emergent phenomena are reported: smooth orientation maps in the first layer, a center–periphery entropy gradient interpreted as a cortical magnification-like pattern, and clustered face/place/tool patches in the final layer. In lesion experiments, removing 50% of units in low-entropy regions retained 90.91% of baseline performance, whereas lesioning high-entropy regions reduced performance to 57.87%. In a new human dataset of spatial biases in object recognition, human accuracy maps correlated with category-specific spatial priors at $p_i^{(l)}$2, while All-TNNs achieved mean $p_i^{(l)}$3 and CNN controls $p_i^{(l)}$4.

A related but lighter mechanism appears in “Biologically Inspired Semantic Lateral Connectivity for Convolutional Neural Networks” (Weidler et al., 2021). There, filters within a layer are coupled along a 1D topographic axis by a fixed Mexican hat lateral connectivity profile, described as a center-surround interaction across the filter domain. The paper states that the stationary distribution of modulated filter activations can be determined analytically, avoiding recurrence, and that the profile orders filters in a sequence resembling the topographic organization of feature selectivity in early visual cortex. Because the provided material does not specify the exact kernel form or numerical results, the main significance lies in the mechanism: topographic organization can emerge not only from auxiliary losses, but also from lateral interaction structure imposed during inference.

These cortical-sheet models share a common premise: global translational weight sharing, the defining bias of standard CNNs, is poorly matched to cortical topography. Topographic alternatives restore locality over units or kernels, then use smoothness, similarity, or wiring-cost surrogates to induce maps.

4. Topographic representations at the input level

In EEG decoding, the topographic problem is not primarily inside the CNN but in the representation delivered to it. “Convolutional Neural Networks with A Topographic Representation Module for EEG-Based Brain-Computer Interfaces” introduces a Topographic Representation Module (TRM) that maps raw EEG $p_i^{(l)}$5 to a 3D topographic map $p_i^{(l)}$6, then uses a convolution block to map back to an output of the same size as the input (Liang et al., 2022). Electrode values are placed at predefined 2D coordinates on a scalp grid; non-electrode cells are zero. The mapped tensor is processed frame-wise by 2D convolutions, and a final convolution collapses the spatial dimensions to recover a $p_i^{(l)}$7 representation. Two variants are defined, TRM-(5,5) and TRM-(3,3), differing only in kernel size.

The significance of the design is architectural compatibility: the output of TRM has the same size as the input, so existing backbones need not be altered. Embedded into DeepConvNet, EEGNet, and ShallowConvNet, TRM improved classification accuracy on both EBDSDD and HGD. On EBDSDD, TRM-(5,5) improved DeepConvNet, EEGNet, and ShallowConvNet by 6.54%, 1.72%, and 2.07%; TRM-(3,3) improved them by 7.76%, 1.71%, and 2.17%. On HGD, TRM-(5,5) improved the same models by 6.05%, 3.02%, and 5.14%, while TRM-(3,3) improved them by 7.61%, 5.06%, and 6.28%. The paper also reports the overhead: train-plus-validation time over 300 epochs increased by 21–158% depending on dataset, backbone, and kernel size.

A cartographic analogue appears in “Deep Convolutional Neural Networks for Map-Type Classification” (Zhou et al., 2018). There, map-type recognition is framed as classification from visual content alone when metadata are absent. The study uses a seven-class deepMap benchmark in which topographic, terrain, and physical maps are merged into one class because of substantial visual overlap, while “The National Map” is kept as a publisher-based class. On the 80/20 split, Inception-ResNet v2 achieved 95–99% overall accuracy, higher than AlexNet, VGG-19, Inception v4, and ResNet v2-152. The paper explicitly notes that “some types of maps, such as the National Map’s topographic maps, are difficult to distinguish without a well-labeled dataset,” underscoring that topographic content often overlaps with publisher style. A plausible implication is that topographic CNNs for cartographic imagery must address both thematic content and style-domain variability.

5. CNNs for physical topography and surface morphology

In materials science, topographic CNNs are used to quantify fracture-surface morphology. “Toward quantitative fractography using convolutional neural networks” employs a U-Net with a pretrained VGG16 encoder to segment SEM images into intergranular, transgranular, and background classes (Tsopanidis et al., 2019). Training used 605 images, validation 105, and test 30, with focused evaluation on 8 fully annotated test images. The optimizer was Adam with initial learning rate $p_i^{(l)}$8, $p_i^{(l)}$9, $$S^{(l)}_{i,j}(y^{(l)}) \triangleq \frac{\sum_{h=1}^{H} \sum_{w=1}^{W} y^{(l)}_{i,h,w}\, y^{(l)}_{j,h,w}} {\left\| \tilde{y}^{(l)}_i \right\|_2 \left\| \tilde{y}^{(l)}_j \right\|_2},$$0, batch size 4, and 40 epochs. On $$S^{(l)}_{i,j}(y^{(l)}) \triangleq \frac{\sum_{h=1}^{H} \sum_{w=1}^{W} y^{(l)}_{i,h,w}\, y^{(l)}_{j,h,w}} {\left\| \tilde{y}^{(l)}_i \right\|_2 \left\| \tilde{y}^{(l)}_j \right\|_2},$$1 test images, after excluding background/void pixels from ground truth, intergranular IoU reached 93.0%, transgranular IoU 87.4%, mean IoU 91.1%, and F-measure 90.7%. Cross-material transfer from $$S^{(l)}_{i,j}(y^{(l)}) \triangleq \frac{\sum_{h=1}^{H} \sum_{w=1}^{W} y^{(l)}_{i,h,w}\, y^{(l)}_{j,h,w}} {\left\| \tilde{y}^{(l)}_i \right\|_2 \left\| \tilde{y}^{(l)}_j \right\|_2},$$2 to $$S^{(l)}_{i,j}(y^{(l)}) \triangleq \frac{\sum_{h=1}^{H} \sum_{w=1}^{W} y^{(l)}_{i,h,w}\, y^{(l)}_{j,h,w}} {\left\| \tilde{y}^{(l)}_i \right\|_2 \left\| \tilde{y}^{(l)}_j \right\|_2},$$3, without retraining, yielded mean total IoU 94% and F-measure 82.4% when void/background were excluded. The quantitative output is the area fraction of each fracture mode, obtained by aggregating hard labels or probabilities over non-void pixels.

In cartography, “Cartographic Relief Shading with Neural Networks” uses a five-level U-Net to render shaded relief from DEMs, learning design principles from Swiss manual shading (Jenny et al., 2020). The network takes normalized DEM tiles, typically $$S^{(l)}_{i,j}(y^{(l)}) \triangleq \frac{\sum_{h=1}^{H} \sum_{w=1}^{W} y^{(l)}_{i,h,w}\, y^{(l)}_{j,h,w}} {\left\| \tilde{y}^{(l)}_i \right\|_2 \left\| \tilde{y}^{(l)}_j \right\|_2},$$4, and predicts a cropped central output tile of $$S^{(l)}_{i,j}(y^{(l)}) \triangleq \frac{\sum_{h=1}^{H} \sum_{w=1}^{W} y^{(l)}_{i,h,w}\, y^{(l)}_{j,h,w}} {\left\| \tilde{y}^{(l)}_i \right\|_2 \left\| \tilde{y}^{(l)}_j \right\|_2},$$5, leaving a 50-pixel contextual border. Dropout is applied at every level with rates 0.075, 0.15, 0.225, 0.375, and 0.525, weights are He-initialized, and training uses Adam with $$S^{(l)}_{i,j}(y^{(l)}) \triangleq \frac{\sum_{h=1}^{H} \sum_{w=1}^{W} y^{(l)}_{i,h,w}\, y^{(l)}_{j,h,w}} {\left\| \tilde{y}^{(l)}_i \right\|_2 \left\| \tilde{y}^{(l)}_j \right\|_2},$$6, $$S^{(l)}_{i,j}(y^{(l)}) \triangleq \frac{\sum_{h=1}^{H} \sum_{w=1}^{W} y^{(l)}_{i,h,w}\, y^{(l)}_{j,h,w}} {\left\| \tilde{y}^{(l)}_i \right\|_2 \left\| \tilde{y}^{(l)}_j \right\|_2},$$7, $$S^{(l)}_{i,j}(y^{(l)}) \triangleq \frac{\sum_{h=1}^{H} \sum_{w=1}^{W} y^{(l)}_{i,h,w}\, y^{(l)}_{j,h,w}} {\left\| \tilde{y}^{(l)}_i \right\|_2 \left\| \tilde{y}^{(l)}_j \right\|_2},$$8, $$S^{(l)}_{i,j}(y^{(l)}) \triangleq \frac{\sum_{h=1}^{H} \sum_{w=1}^{W} y^{(l)}_{i,h,w}\, y^{(l)}_{j,h,w}} {\left\| \tilde{y}^{(l)}_i \right\|_2 \left\| \tilde{y}^{(l)}_j \right\|_2},$$9, batch size 8, and 1,000–5,000 epochs. A $$\tau^{(l)}_{i,j} \triangleq \frac{1}{d^{(l)}(i,j)+1},$$0 DEM can be rendered in about 6 seconds on a 2019 laptop with AMD Radeon Pro 5300M GPU. In an expert study with 18 relief-shading specialists, CNN shadings were rated 15% poor/fair, 32% good, 43% very good, and 10% excellent, and swisstopo decided to use the approach to extend coverage of manual shaded relief at 1:200,000 and 1:500,000 scales.

For coastal prediction, “TopoFormer: Integrating Transformers and ConvLSTMs for Coastal Topography Prediction” addresses extrapolation from MLWN to MLWS using beach profiles represented as ordered elevation–chainage pairs (Munian et al., 11 Jan 2025). After preprocessing, 1,366 usable profiles remained across 33 locations. Inputs concatenate 80 elevation samples down to MLWN with 100 chainage samples, and the output is 20 elevation predictions for the MLWN→MLWS segment. TopoFormer stacks six transformer blocks, each containing multi-head self-attention, layer normalization, and four ConvLSTM layers, followed by a two-layer MLP and a dense output layer. The model has 761K trainable parameters and achieved MAE = 0.021 and RMSE = 0.026 on the test set, outperforming DenseNet, 1D/2D CNNs, LSTM, biLSTM, and ConvLSTM. A plausible implication is that, for physical topography, convolution alone may be insufficient when long-range dependencies along the profile are critical.

In robotics, “Neural Scene Representation for Locomotion on Structured Terrain” instantiates a topographic CNN as a sparse 4D fully convolutional network on point clouds over space and time (Hoeller et al., 2022). Multi-camera depth is fused into a robot-centric $$\tau^{(l)}_{i,j} \triangleq \frac{1}{d^{(l)}(i,j)+1},$$1 voxel grid covering a $$\tau^{(l)}_{i,j} \triangleq \frac{1}{d^{(l)}(i,j)+1},$$2 cube, and the network uses the current measurement together with the previous reconstruction transformed into the current frame. On structured terrains, the method achieved on stairs precision 86.0%, recall 89.9%, F1 88.9%, and MAE 0.8 cm, compared with Elevation Mapping F1 76.3%, MAE 2.1 cm, and Voxblox F1 73.5%, MAE 1.6 cm. On a Jetson Xavier, inference ran in about 70 ms per frame, with the full ROS pipeline at about 6 Hz.

An exploratory remote-sensing counterpart appears in “Towards Monocular Digital Elevation Model (DEM) Estimation by Convolutional Neural Networks - Application on Synthetic Aperture Radar Images” (Costante et al., 2018). A fully convolutional encoder–decoder maps single-pass SAR magnitude and phase channels to DEM estimates in radar coordinates. Using 4,000 × 4,000 sliding windows downsampled to $$\tau^{(l)}_{i,j} \triangleq \frac{1}{d^{(l)}(i,j)+1},$$3, Adam at learning rate 0.001, batch size 128, and 500 epochs, the paper reports average test RMSE of 105.28 m for Alps, 74.46 m for California, and 43.45 m for Tucson. The authors explicitly characterize the approach as exploratory and note that sharp crests and high-relief details are smoothed under the RMSE objective.

Across these applications, “topographic CNN” refers less to a fixed architectural schema than to CNNs tailored to scalar or morphological structure defined over surfaces, profiles, or volumes.

6. Local connectivity prediction and topological reconstruction

A related line of work addresses topology rather than scalar topography, but it is methodologically close because it converts local spatial predictions into global structured maps. “Iterative Deep Learning for Road Topology Extraction” uses two components: a global dense semantic model producing per-pixel road probability maps, and a local connectivity CNN that predicts which border points of a patch are connected to the center (Ventura et al., 2018). For a patch centered at $$\tau^{(l)}_{i,j} \triangleq \frac{1}{d^{(l)}(i,j)+1},$$4 with border points $$\tau^{(l)}_{i,j} \triangleq \frac{1}{d^{(l)}(i,j)+1},$$5, the CNN outputs $$\tau^{(l)}_{i,j} \triangleq \frac{1}{d^{(l)}(i,j)+1},$$6 and is trained with weighted binary cross-entropy:

$$\tau^{(l)}_{i,j} \triangleq \frac{1}{d^{(l)}(i,j)+1},$$7

Inference traces the network iteratively: from seed points, the model expands to border points with $$\tau^{(l)}_{i,j} \triangleq \frac{1}{d^{(l)}(i,j)+1},$$8, adds corresponding edges, snaps nodes to the road centerline if desired, merges nearby nodes, and stops when the queue is empty or all high-confidence road pixels lie near visited nodes. The method’s complexity is $$\tau^{(l)}_{i,j} \triangleq \frac{1}{d^{(l)}(i,j)+1},$$9.

The paper positions this against pure segmentation plus skeletonization. Standard FCN/U-Net segmentation captures “where” road pixels lie, but not “how things connect”; the patch-based connectivity predictor instead encodes explicit local topology and reconstructs a graph by chaining local decisions. The same paradigm generalizes to retinal vessels, where continuity through crossings and bifurcations is crucial. This suggests an important conceptual boundary within the broader topographic-CNN literature: some systems predict scalar surfaces, while others infer graph structure embedded in space.

7. Empirical themes, limitations, and research directions

Several themes recur across these papers. First, topographic inductive bias often preserves or only slightly reduces benchmark accuracy while improving structural properties such as pruning resistance, robustness to corruption, or spatial localization. This is explicit in channel-lattice topographic loss models, which match baseline accuracy across four datasets while improving robustness to pruning, and in WS-constrained topographic layers, which trade small accuracy losses at strong $d^{(l)}(i,j)=\|p_i^{(l)}-p_j^{(l)}\|_2$0 for smoother and more robust representations (Poli et al., 2022, Truong et al., 31 Jul 2025). Second, the specific implementation matters: activation-based and weight-based spatial constraints are not interchangeable, and locally connected topographic architectures differ qualitatively from convolutional architectures with added regularizers.

Third, computational cost is a recurring concern. Topographic loss based on full channel-pair similarities requires $d^{(l)}(i,j)=\|p_i^{(l)}-p_j^{(l)}\|_2$1 memory and compute per organized layer; TRM adds 21–158% training time; locally connected All-TNNs sacrifice the efficiency of weight sharing; and sparse 4D terrain reconstruction requires specialized sparse tensor implementations (Poli et al., 2022, Liang et al., 2022, Lu et al., 2023, Hoeller et al., 2022). These costs explain why many papers apply topography only to selected layers or use compact grids.

Fourth, several studies identify domain-specific ambiguities that topographic inductive bias alone does not resolve. In cartographic classification, confusion persists between the merged topographic/terrain/physical class and The National Map because of shared contour, hillshade, and color conventions. In road tracing, overpasses, dense urban layouts, and large occlusions remain failure modes, and performance depends sensitively on patch size, $d^{(l)}(i,j)=\|p_i^{(l)}-p_j^{(l)}\|_2$2, and threshold $d^{(l)}(i,j)=\|p_i^{(l)}-p_j^{(l)}\|_2$3. In monocular SAR DEM estimation, the problem remains ill-posed because a single SLC image lacks an explicit height observable. In relief shading, models trained at coarse cell sizes may fail on very fine DEMs, especially in flatter areas (Zhou et al., 2018, Ventura et al., 2018, Costante et al., 2018, Jenny et al., 2020).

A broader synthesis is that topographic CNNs are most effective when the learning problem has a genuine spatial organization that standard convolution either ignores or over-symmetrizes. In some cases the remedy is to spatialize channels, in others to spatialize inputs, and in others to predict explicitly topographic outputs. The literature therefore supports a family resemblance rather than a single definition: topographic CNNs are convolutional or convolution-derived models that encode, exploit, or reconstruct structured spatial organization beyond ordinary translational equivariance.