Implicit Neural Spatial Representations (INSRs)

• Implicit Neural Spatial Representations (INSRs) are continuous neural mappings that embed spatial coordinates into latent spaces and use deep architectures to model complex signals.
• They extend classical implicit representations by incorporating hybrid MLP-convolution blocks and compression-expansion bottlenecks, enabling both pixel-level fitting and dataset-level learning.
• Empirical results demonstrate significant gains in image fitting, classification, detection, and segmentation, underscoring the versatility and scalability of INSR frameworks.

Implicit Neural Spatial Representations (INSRs) generalize the paradigm of continuous coordinate-to-signal mapping via neural networks, elevating INRs from fitting individual signals to modeling structured objects or datasets. By embedding spatial coordinates or higher-order elements into latent spaces and parameterizing the mapping with deep neural architectures, INSRs support both low-level and high-level vision tasks, including image fitting, classification, object detection, and segmentation. This entry synthesizes key advances in the definition, architecture, and empirical evaluation of INSRs, with focus on recent developments such as the Implicit Neural Representation Network (INRN) (Song et al., 2022).

1. Theoretical Foundations of INSRs

Classical implicit neural representations are functions, typically parameterized as $f:\mathbb{R}^2 \to \mathbb{R}^3$, that map spatial image coordinates $(x, y)$ to RGB values. However, this formulation does not generalize across datasets or high-level tasks. The INSR framework extends this notion by treating the representation as a network $$f_\theta:\mathbb{R}^E \to \mathbb{R}^{\mathrm{Output}}$$, where sub-elements $\mathbf{x}_i$ (pixels, images, etc.) are first embedded in $\mathbb{R}^E$ via a fixed or learnable embedding $\Phi$. The network thus models arbitrary structured objects $\mathcal{D} = \{(\mathbf{x}_i, f(\mathbf{x}_i))\}$:

$$\hat{f}_\theta:\underbrace{\Phi(\mathbf{x})}_{\in \mathbb R^E} \longrightarrow \underbrace{f(\mathbf{x})}_{\in \mathbb R^{\rm Output}}$$

This abstraction accommodates both pixel-wise fitting (low-level) and instance-wise or dataset-wise parameterizations (high-level tasks).

2. INRe Basic Block: Hybrid MLP–Convolutional Architecture

The INRe block, central to INRN, combines convolutional and multilayer perceptron pathways to inject spatial inductive bias and enable parameter control. Each block operates as follows:

• Hybrid structure: Alternates $1\times 1$ convolution and a two-layer MLP. Given input features $$\mathbf{z}^{(\mathrm{in})}\in\mathbb{R}^{H\times W\times C_\mathrm{in}}$$, apply

$$\mathbf{u} = \mathrm{Conv}_{1\times1}(\mathbf{z}^{(\mathrm{in})}) \in \mathbb{R}^{H \times W \times C'}$$

Flatten spatial dimensions and process features through a GELU-activated MLP:

$$\mathbf{v} = \sigma\left(W_2\,\mathrm{GELU}(W_1\,\mathrm{vec}(\mathbf{u}) + b_1) + b_2\right)$$

Reshape the output as necessary.

• Compression-expansion bottleneck: The MLP compresses from $C' \to rC'$ channels (compression), then expands to $C_\mathrm{out}$, using:

$$\mathbf{h}_1 = \mathrm{GELU}(W_c\mathbf{u} + b_c),\quad \mathbf{h}_2 = W_e\mathbf{h}_1 + b_e$$

Channel bottlenecks reduce parameter count and empirical error.

• Activation: GELU ($\mathrm{GELU}(x) = x\cdot\Phi(x)$) is preferred over ReLU for smoother gradients and training stability.

3. Deep INRN: Stacking Strategies and Loss Functions

INRN utilizes two stacking paradigms tailored for different task regimes:

• Single-stage INRN: For low-level tasks, stack $L$ INRe blocks, outputting a final image $$\hat{Y} = f_\theta(X) \in \mathbb{R}^{H \times W \times 3}$$. The loss is a weighted sum of MSE and SSIM:

$$\mathcal{L} = \alpha\frac{1}{N}\|f_\theta(X) - Y\|_2^2 + (1-\alpha)[1 - \mathrm{SSIM}(f_\theta(X), Y)]$$

• Multi-stage INRN: For high-level tasks, divide the network into $S$ stages. At each stage $i$, align the INRN feature $\mathbf{O}_s^i$ to a teacher network's output $\mathbf{O}_t^i$ using a $1\times 1$ convolutional aligner:

$$\mathcal{L}_{ms} = \sum_{i=1}^S \mathrm{MSE}\left(\mathcal{T}_s^i(\mathbf{O}_s^i),\,\mathcal{T}_t^i(\mathbf{O}_t^i)\right)$$

The total loss for classification/detection is:

$$\mathcal{L}_\mathrm{final} = \lambda_1\,\mathcal{L}_{\rm CE} + \lambda_2\,\mathcal{L}_{ms}$$

4. Comparative Empirical Performance

INRN exhibits measurable performance improvements in both image fitting and high-level vision tasks.

Task	Baseline	INRN Result	Gain
Low-level image fitting	Pure MLP: 25.39 PSNR	INRe: 32.13 PSNR	+6.74 dB
CIFAR-100 classification	KD: 70.66%, Ours:71.06%	+0.4 pt
ImageNet classification	CRD: 77.17%, Ours:76.70%	Comparable
COCO object detection	R50: 37.93 mAP	R50+INRN: 39.08 mAP	+1.15 mAP
Instance segmentation	R50: 35.24 mAP	R50+INRN: 36.52 mAP	+1.28 mAP

Stage allocation ablations also show improvements: for ImageNet, allocating more blocks as [2,3,5,2] yields 68.43% vs 66.74% (+1.7 pt); COCO detection shows +5.5 mAP with nonuniform block allocation.

5. Architectural Implications for Representation Capacity

Conventional INRs (e.g., SIREN, LIIF) are shallow MLPs that only fit single signals. INRN fundamentally extends the representation power by:

• Hybridizing spatial convolution and MLP channels within each block, injecting spatial bias and controlling capacity.
• Compression-expansion bottlenecks mitigate over-parameterization, reducing risk of overfitting and instability.
• GELU activations ensure robust optimization and smoother loss landscapes.
• Deep stacking (multi-stage) permits dataset-level parameterizations for complex vision tasks; i.e., $\mathbf{x}_i$ as entire images.

6. Broader Impact and Future Directions

By formalizing INSRs through INRN, the field achieves scalable implicit representations suitable for high-level cognitive vision problems. Key outcomes include:

• Extension of INRs from pixel-level fitting to dataset-level learning (classification, detection, segmentation).
• Consistent empirical gains across tasks and architectures.
• Architectural modularity (hybrid Conv-MLP, bottlenecking, spatial bias) as transferable design principles.

These advances highlight that implicit neural spatial representations, when properly equipped with deep, hybridized, and bottlenecked designs, can serve as general-purpose, high-capacity models for structured vision learning, challenging the notion that INRs are suitable only for low-level signal fitting (Song et al., 2022).