JA-SIREN eliminates random initialization for sinusoidal networks, boosting PSNR by 21 dB

Implicit neural representations (INRs) have become a powerful tool for representing signals as neural networks, but they suffer from a major flaw: stochastic initialization. Current approaches like SIREN rely on random seeds, leading to inconsistent performance across runs—variations exceeding 2.5 dB (78%) in image regression are common. For scientific computing and simulation, where reproducibility is mandatory, this unpredictability is unacceptable. Researchers from Carnegie Mellon University and the University of Waterloo have now introduced JA-SIREN (Jacobi-Anger Sinusoidal Representation Network), a deterministic initialization scheme that eliminates randomness entirely. The method leverages classical spectral analysis: it first computes the Discrete Sine Transform (DST) of the target signal, then uses the Jacobi-Anger expansion to derive closed-form weights for a two-layer sinusoidal MLP. This analytically matches the network's initial spectral response to the signal without requiring any random seed or additional hyperparameter tuning.

The results are striking. On the standard Kodak dataset, JA-SIREN achieves a mean PSNR of 67.18 dB—a 21.30 dB improvement over the best existing baseline. More importantly, it exhibits zero run-to-run variance, meaning every training session yields identical results. This deterministic behavior is a breakthrough for applications in scientific computing, inverse problems, and simulation, where consistent outputs are essential. The paper also demonstrates that spectrally-informed initialization is fundamentally more effective than stochastic approaches for sinusoidal INRs. By grounding neural network initialization in well-understood signal processing theory, JA-SIREN not only boosts performance but also removes one of the key hurdles to deploying INRs in production environments. The work is available on arXiv and has been submitted to a top computer vision venue.