Research map / Architectures and efficiency / RNNs and normalization
EVOS: Efficient Implicit Neural Training via EVOlutionary Selector
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes EVOS, an evolutionary sample-selection paradigm (sparse fitness evaluation, frequency-guided crossover, augmented unbiased mutation) that trains implicit neural representations on a sparse subset of coordinates per iteration, cutting training time by roughly 48-66%.
Problem
Implicit neural representation (INR) training is computationally expensive because every iteration forwards all coordinates through the MLP, and spectral bias further slows fitting of high-frequency detail.
Approach
Treats coordinates as individuals in an evolutionary process: sparse, cached fitness evaluation selects survivors, frequency-guided crossover combines low- and high-frequency parents, and unbiased mutation adds random non-selected points, so the network forwards only a sparse subset each iteration, with a cross-frequency loss to supervise it.
Key results
- EVOS reaches 35 dB PSNR in 66.46% less time than standard training (29.73 s vs 88.64 s) on DIV2K with a 3x256 SIREN.
- Under fixed 5k iterations, EVOS achieves 37.81 PSNR versus 37.10 for standard training while cutting time from 180.45 s to 97.39 s.
- Reported 48%-66% time reduction across target-PSNR settings, with gains consistent across PEMLP, SIREN, WIRE, GAUSS and FINER backbones.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use EVOS instead of Standard training | SIREN 3×256 · DIV2K (processed), fixed 5k iterations | PSNR: 37.10 -> 37.81 | Helped |
| use EVOS instead of Standard training | SIREN 3×256 · DIV2K (processed), fixed 5k iterations | time: 180.45 -> 97.39 | Helped |
| use EVOS instead of Standard training | SIREN 3×256 · DIV2K (processed), target PSNR 25 dB | time: 9.04 -> 4.73 | Helped |
| use EVOS instead of Standard training | SIREN 3×256 · DIV2K (processed), target PSNR 30 dB | time: 30.75 -> 12.48 | Helped |
| use EVOS instead of Soft Mining | SIREN 3×256 · DIV2K (processed), target PSNR 35 dB | time: 84.81 -> 29.73 | Helped |
| add sparse fitness evaluation (ablation) | EVOS · DIV2K (processed), ablation study | PSNR: 35.41 -> 37.81 | Helped |
| use θ = 0 instead of the default schedule (ablation) | EVOS · DIV2K (processed), ablation study | PSNR: 37.81 -> 37.54 | Hurt |
| set τ = 1 instead of the default schedule (ablation) | EVOS · DIV2K (processed), ablation study | time: 97.39 -> 142.39 | Hurt |
| set τ = 10 instead of the default schedule (ablation) | EVOS · DIV2K (processed), ablation study | PSNR: 37.81 -> 36.37 | Hurt |
| add frequency-guided crossover (ablation) | EVOS · DIV2K (processed), ablation study | PSNR: 37.04 -> 37.81 | Helped |
| add cross-frequency supervision (ablation) | EVOS · DIV2K (processed), ablation study | PSNR: 37.49 -> 37.81 | Helped |
| add augmented unbiased mutation (ablation) | EVOS · DIV2K (processed), ablation study | PSNR: 34.94 -> 37.81 | Helped |
Surprising findings
- Sparsified sample selection improves per-iteration fitting quality over full-data training under identical iteration counts (EVOS w/o CFS reaches 31.64 PSNR at 1k iterations vs 31.06 standard), which the authors attribute to sample-weighting effects.
- Increasing evaluation frequency (smaller tau) did not improve performance despite higher cost, which the authors attribute to training instability.
Limitations
- Gains are reported mainly on 2D image fitting (DIV2K); other modalities only in supplementary material.
- Relies on periodic cached fitness evaluations, and overly frequent evaluation reduced stability.
- Some backbones show mixed results (e.g., WIRE and FINER at 5k iterations show small or negative PSNR changes).
Experimental setup
Datasets: DIV2K
Benchmarks: DIV2K
Metrics: PSNR SSIM LPIPS training time (sec)
Baselines: Uniform sampling EGRA Expansive Supervision Soft Mining INT Standard full-coordinate training
Builds on: SIREN INT (nonparametric teaching) Expansive Supervision EGRA
Closest papers by meaning
- Guided evolutionary strategies: augmenting random search with surrogate gradients (ICML 2019 · similarity 0.87)
- NeuroEvoBench: Benchmarking Evolutionary Optimizers for Deep Learning Applications (NeurIPS 2023 · similarity 0.87)
- EG-NAS: Neural Architecture Search with Fast Evolutionary Exploration (AAAI 2024 · similarity 0.85)
- RoSA: Rotational Sparse Adaptation for Memory-Efficient Fine-Tuning (NeurIPS 2026 · similarity 0.85)
- Runtime Analysis of Evolutionary NAS for Multiclass Classification (ICML 2025 · similarity 0.84)
- MADA: Meta-Adaptive Optimizers Through Hyper-Gradient Descent (ICML 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.