- The paper introduces CHOIR, which replaces deep compositional periodic INRs with a hybrid additive superposition to stabilize optimization.
- It employs Perceptual Spectrum Calibration to assign log-uniform frequency bases, effectively mitigating spectrum bias in natural signals.
- CHOIR demonstrates superior performance in high-dimensional signal recovery tasks, including image synthesis and multi-modal data reconstruction.
Calibrated Harmonic Overlaid Implicit Neural Representations for Multi-Dimensional Data
Introduction and Motivation
Implicit neural representations (INRs) parameterize signals as coordinate-wise continuous functions, becoming foundational for multi-dimensional signal recovery tasks. The usage of periodic activation functions such as Sine in SIREN [38] and its derivatives allows these networks to recover high-frequency information efficiently. However, prevailing approaches are limited by two core architectural drawbacks: first, the heavy reliance on deep function composition contradicts the additive superposition principle underlying harmonic signal representations, causing notable optimization instability as depth increases; second, the fixed or heuristically initialized frequency bases employed induce significant spectrum bias, mismatched with the scale-free (1/f) spectral statistics characteristic of natural signals.
The paper introduces CHOIR (Calibrated Harmonic Overlaid Implicit Neural Representation), addressing these challenges through Coordinated Harmonic Superposition (CHS) for architectural stability and Perceptual Spectrum Calibration (PSC) for spectrum bias mitigation. These modifications systematically enable deeper, more robust periodic INRs that better model multi-dimensional signals.
Traditional periodic INRs construct functions through nested compositions of periodic non-linearities. As depth increases, the repeated application of these activations leads to Jacobian products in backpropagation, directly causing vanishing/exploding gradients and optimization failure. Furthermore, the composition-based architecture lacks an explicit mapping from learnable coefficients to basis functions, disrupting harmonious representation and leaving high-frequency details under-optimized or overfit.
CHOIR replaces this composition paradigm with a hybrid composition-superposition architecture. Each network layer comprises harmonic modules whose outputs are scaled by learnable, initially zeroed scalars and added to the running intermediate representation. This results in explicit additive superposition (Eq. 4), mapping directly onto generalized Fourier series and ensuring stable optimization. The CHS mechanism invokes an implicit curriculum learning effect, as harmonic terms activate progressively when their contributions align with the global descent direction during gradient-based training. This not only improves stability but hierarchically unfolds the representation capacity as dictated by the optimization landscape and data complexity.
Perceptual Spectrum Calibration
Spectrum bias is endemic in coordinate INRs, as fixed frequency bases yield updates primarily benefiting low-frequency content due to the power-law decay of natural spectra. CHOIR's PSC directly incorporates a log-uniform frequency assignment in the harmonic modules. Frequencies per layer are geometrically distributed from a minimal fundamental to an upper bound, scaled below the Nyquist rate to avoid training instabilities caused by high-frequency neuron dominance. This frequency allocation matches the 1/f energy profile typical of natural signals.
To regularize the influence of high-frequency harmonics, amplitudes are initialized (and subsequently adapted) according to the root of the expected power spectrum, enforcing a balance where gradients from all frequency bands are equally represented at initialization. The spectral decay parameter is global and trainable, allowing the network to re-tune amplitude scaling in accordance with the true spectrum of the task data. This dual frequency and amplitude assignment both constrains and adapts the solution space, enhancing the optimization properties and generalization of the network.
Experimental Results
Extensive quantitative and qualitative evaluations position CHOIR as consistently outperforming alternatives on signal fitting, view synthesis, and various multi-modal data recovery tasks. For 2D image fitting on Kodak data, CHOIR surpasses SIREN, FINER, FreSh, and others by large margins in PSNR, SSIM, and LPIPS, recovering high-frequency detail unattainable by these competitors.
On the NeRF Blender 5D scene data, CHOIR achieves superior performance in both metrics and visual sharpness of synthesized views, highlighting improved generalization. Across standard multi-dimensional datasets—hyperspectral (HSI), multispectral (MSI), RGB, and videos—CHOIR demonstrates robust handling in the face of extreme missing data and complex noise, outperforming tensor factorization-based methods (LRTFR, DRO-TFF, CRNL) and periodic INR baselines (SIREN, FINER) in all designed test scenarios.
Particularly notable is the stable scaling of CHOIR: whereas other periodic INRs suffer from degradation as depth increases, CHOIR's PSNR consistently improves with added layers, validating the architectural superiority of CHS. Ablation studies confirm the necessity of both CHS and PSC for optimal recovery; either component alone yields lower reconstruction quality.
From a computational efficiency perspective, CHOIR achieves state-of-the-art fidelity with substantially fewer parameters and lower runtime compared to high-performing prior work (e.g., CRNL). Neural Tangent Kernel (NTK) analysis supports the theoretical claim that CHOIR's initialization provides stronger spatial independence, enabling data-adaptive, decoupled optimization of high-frequency content.
Implications and Future Directions
Practically, CHOIR's principled architectural and spectral calibration advances the methodology for high-fidelity recovery of multi-modal signals, with broad relevance in computational imaging, remote sensing, and scientific data analysis. Theoretically, its design highlights the importance of harmonic alignment—both in network structure and inductive priors—when modeling signals with scale-free spectra. The superposition-based approach disentangles the representation bottlenecks that have previously limited the expressiveness and scalability of periodic INRs.
For future development, extension towards irregular or non-Euclidean domain data (e.g., point clouds, event streams, spatial transcriptomics) is natural. The demonstrated stability and adaptability provide a foundation for deploying deeper, broader implicit representations in practical and scientific systems where high-frequency fidelity and generalization are crucial.
Conclusion
CHOIR overcomes longstanding obstacles in deep periodic INR optimization and spectrum bias by unifying harmonic superposition with calibrated frequency allocation. The architecture enables deep, stable, and efficient INRs, advancing the frontier of multi-dimensional signal recovery with superior empirical performance, theoretical soundness, and promising generalization (2606.26763).
