Research map / Architectures and efficiency / RNNs and normalization
End-to-End Implicit Neural Representations for Classification
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.
Meta-learns a SIREN initialization and per-parameter learning-rate schedule end-to-end with a Transformer classifier over SIREN weights (Meta Weight Transformer), achieving SOTA SIREN classification on MNIST, Fashion-MNIST, CIFAR-10, and first INR classification on ImageNet-1K.
Problem
Classifying images via their implicit neural representation (SIREN) weights underperforms pixel-based methods, because weight-space symmetries make parameters hard for downstream models to interpret and existing pipelines fit INRs independently of the classifier.
Approach
Back-propagate the classification loss through a few-step, meta-learned SIREN fitting process so the shared initialization and learning-rate schedule are optimized for both reconstruction and classification; then apply a standard Transformer (no symmetry equivariance) to the SIREN weight tokens, using differences from the base initialization.
Key results
- On CIFAR-10 without augmentation, MWT-L reaches 59.57% accuracy versus 38.8% for the prior SOTA ScaleGMN-B.
- With augmentation on CIFAR-10, the method reaches 64.7%, improving over inr2array's 63.4%.
- On full ImageNet-1K, MWT-L achieves 24.11% accuracy (23.6% in abstract) with a subsampling rate of 0.01.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add end-to-end classifier gradients to guide SIREN meta-learning (ablation) | MWT versus WT · CIFAR-10 validation | accuracy: 38.98 -> 55.79 | Helped |
| Use spatial augmentations and a 20-layer Transformer | MWT-L · CIFAR-10 | accuracy: 63.4% -> 64.7% | Helped |
| Set task influence wtask=0.01 rather than 0 (ablation) | MWT · CIFAR-10 validation | accuracy: 42.24 -> 55.79 | Helped |
| Increase task influence from wtask=0.01 to wtask=0.1 (ablation) | MWT · CIFAR-10 validation | accuracy: 55.79 -> 54.00 | Hurt |
| Increase the number of inner-loop steps from 1 to 6 (ablation) | MWT · CIFAR-10 validation | accuracy: 49.70 -> 55.79 | Helped |
| Increase inner-loop steps from 6 to 8 (ablation) | MWT · CIFAR-10 validation | accuracy: 55.79 -> 55.52 | Hurt |
| Increase SIREN width from 128 to 256 (ablation) | MWT · CIFAR-10 validation | accuracy: 55.79 -> 57.19 | Helped |
| Increase Transformer depth from 10 to 20 (ablation) | MWT · CIFAR-10 validation | accuracy: 55.79 -> 56.82 | Helped |
| Share Meta-SGD learning rates (ablation) | MWT · CIFAR-10 validation | accuracy: 55.79 -> 54.38 | Hurt |
| Reduce subsampling rate from s=1.0 to s=0.1 (ablation) | MWT · CIFAR-10 validation | accuracy: 55.79 -> 54.21 | Hurt |
| Use MWT-L with s=0.02 rather than MWT with s=0.05 (ablation) | MWT and MWT-L · Imagenette, with augmentation | accuracy: 57.27 ± 0.38 -> 60.75 ± 0.37 | Helped |
| Increase task influence from wtask=0.0001 to wtask=0.01 (ablation) | MWT-L · ImageNet-1K | accuracy: 19.21% -> 24.11% | Helped |
Surprising findings
- Using only a random 5% of pixels per inner-loop step gave the best small-model Imagenette accuracy (57.27%), with little degradation from subsampling.
- Increasing the classifier's influence on meta-learning beyond a point lowers both reconstruction quality and classification accuracy.
Limitations
- Authors are uncertain whether initialization and learning-rate schedule alone impose sufficient structure on SIREN parameters.
- Reconstruction quality is considered sufficient but may be improvable, especially on high-resolution datasets.
- Higher inner-loop step counts are computationally expensive, a known MAML limitation.
Experimental setup
Datasets: MNIST Fashion-MNIST CIFAR-10 Imagenette ImageNet-1K ModelNet40
Benchmarks: MNIST Fashion-MNIST CIFAR-10 Imagenette ImageNet-1K ModelNet40
Metrics: classification accuracy PSNR
Baselines: MLP INR2Vec DWS NFN NG-GNN ScaleGMN ScaleGMN-B WT NFN-NP NFN-HNP INR2ARRAY
Builds on: MAML Meta-SGD SIREN Neural Functional Transformers DWS NFN
Closest papers by meaning
- MIRE: Matched Implicit Neural Representations (CVPR 2025 · similarity 0.87)
- Learning Transferable Features for Implicit Neural Representations (NeurIPS 2024 · similarity 0.87)
- Towards Croppable Implicit Neural Representations (NeurIPS 2024 · similarity 0.86)
- Implicit Representations via Operator Learning (ICML 2024 · similarity 0.86)
- VI^3NR: Variance Informed Initialization for Implicit Neural Representations (CVPR 2025 · similarity 0.86)
- SINR: Sparsity Driven Compressed Implicit Neural Representations (CVPR 2025 · similarity 0.85)
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.