Research map / Architectures and efficiency / Pruning and efficient tuning
A Simple Cache Model for Image Recognition
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add a cache component using features from layers close to, but not including, the output layer | Pretrained image classifiers · CIFAR-10+, CIFAR-100+, ImageNet | error rate: 8.33 -> 7.58; 32.66 -> 29.18; 30.98 -> 30.42 | Helped |
| Use cache predictions alone instead of combining them with network predictions (ablation) | Cache models · CIFAR-10+, CIFAR-100+, ImageNet | error rate: 7.01 -> 11.13; 29.36 -> 38.40; 30.42 -> 41.24 | Hurt |
| Use layers close to the output, excluding the output layer, as cache keys (ablation) | Cache component · CIFAR-10, CIFAR-100, ImageNet | test accuracy | Helped |
| Use input-layer features as cache keys (ablation) | Cache component · CIFAR-10, CIFAR-100 | test performance | Hurt |
| Concatenate features from multiple layers to form cache keys (ablation) | Cache component · Image recognition benchmarks | classification performance | Helped |
| Increase cache size (ablation) | ResNet32 cache models · CIFAR-10, CIFAR-100 | test accuracy | Helped |
| Add a Cache3 component | ResNet32 · CIFAR-10 adversarial images generated by FGSM | classification accuracy: 5.2 -> 72.8 | Helped |
| Use cache predictions alone (ablation) | ResNet32-Cache3 · CIFAR-10 adversarial images generated by I-FGSM, SP, and GB | classification accuracy: 68.4 -> 83.2; 58.8 -> 68.6; 56.0 -> 61.3 | Helped |
| Add a Cache3 component | ResNet32 · CIFAR-10 direct white-box attacks | mean minimum perturbation size: 0.064 -> 0.083; 0.014 -> 0.020; 0.044 -> 0.149; 0.224 -> 0.299 | Helped |
| Use cache predictions alone (ablation) | ResNet32-Cache3 · CIFAR-10 direct white-box attacks | attack success | Helped |
| Add a Cache3 component to the transfer model | ResNet20 · CIFAR-10 adversarial examples transferred from ResNet32 | classification accuracy: 77.6 -> 80.8; 89.2 -> 91.6; 70.6 -> 80.4; 54.8 -> 59.7 | Helped |
| Add a cache component | ResNet32 · CIFAR-10 test points | Jacobian norm | Mixed |
Closest papers by meaning
- VL-Cache: Sparsity and Modality-Aware KV Cache Compression for Vision-Language Model Inference Acceleration (ICLR 2025 · similarity 0.86)
- Towards flexible perception with visual memory (ICML 2025 · similarity 0.85)
- $\text{D}_{2}\text{O}$: Dynamic Discriminative Operations for Efficient Long-Context Inference of Large Language Models (ICLR 2025 · similarity 0.84)
- Dual Memory Networks: A Versatile Adaptation Approach for Vision-Language Models (CVPR 2024 · similarity 0.84)
- AsCAN: Asymmetric Convolution-Attention Networks for Efficient Recognition and Generation (NeurIPS 2024 · similarity 0.84)
- AttentionPredictor: Temporal Patterns Matter for KV Cache Compression (NeurIPS 2025 · 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.