Optimizing Rank for High-Fidelity Implicit Neural Representations
Researchers have demonstrated that the perceived inability of standard Multi-Layer Perceptrons (MLPs) to represent high-frequency content in Implicit Neural Representations (INRs) is not an architectural limitation. Instead, they propose that stable rank degradation during training is the primary cause. By regulating the network's rank during training, even simple MLP architectures can achieve significantly higher fidelity, showing improvements of up to 9 dB PSNR across various domains like image synthesis and medical imaging. AI
IMPACT This research suggests that existing MLP architectures may be more capable for complex signal representation than previously thought, potentially reducing the need for specialized architectures in certain INR applications.