Research map / Learning theory and algorithms / Prediction and decision losses
Approximation Algorithms for Cascading Prediction Models
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 |
|---|---|---|---|
| Use a confidence-based greedy cascade of pre-trained models | ImageNet models relative to NasNet · ImageNet validation | average multiplications: 1.5x reduction with no loss in accuracy | Helped |
| Use a confidence-based greedy cascade of pre-trained models | ImageNet models relative to Inception-v4 · ImageNet validation | average multiplications: 1.8x cost reduction | Helped |
| Use a different cascade operating point relative to Inception-v4 | ImageNet models · ImageNet validation | accuracy: 1.2% increase | Helped |
| Use a confidence-based greedy cascade of pre-trained models | ImageNet models relative to largest MobileNet · ImageNet validation | average multiplications: 2x cost reduction | Helped |
| Use a confidence-based greedy cascade of pre-trained models | ImageNet models relative to largest MobileNet · ImageNet validation | accuracy: 0.5% accuracy gain | Helped |
| Cascade quantized models while optimizing memory-read cost | Pool of 368 quantized ImageNet models · ImageNet validation | average memory I/O: up to a factor of 6 with no loss in accuracy | Helped |
| Optimize memory-read cost using a pool of quantized models | ImageNet cascade · ImageNet validation | average multiplications | Helped |
| Threshold logit gap to abstain on uncertain examples | Inception-v3 · ImageNet validation | top-1 accuracy on retained examples: 77% -> over 95% | Helped |
| Use the learned accuracy model instead of directly thresholding logit gap | Inception-v3 abstention rule · ImageNet validation | accuracy-response tradeoff | No clear effect |
| Use lower-resolution, cheaper models at earlier cascade stages | MobileNet cascade · ImageNet validation | multiplications per stage: 49M; 77M; 162M; 150M; 569M | Helped |
| Allocate easier examples to earlier cascade stages | MobileNet cascade · ImageNet validation | stage accuracy: 88%; 73%; 62%; 45%; 45% | Mixed |
Closest papers by meaning
- Model Cascades with Provable Per-Class Quality (NeurIPS 2026 · similarity 0.85)
- Learning to Cascade: Confidence Calibration for Improving the Accuracy and Computational Cost of Cascade Inference Systems (AAAI 2021 · similarity 0.84)
- Gatekeeper: Improving Model Cascades Through Confidence Tuning (NeurIPS 2025 · similarity 0.84)
- Non-vacuous Generalization Bounds at the ImageNet Scale: a PAC-Bayesian Compression Approach (ICLR 2019 · similarity 0.84)
- Stochastic Amortization: A Unified Approach to Accelerate Feature and Data Attribution (NeurIPS 2024 · similarity 0.83)
- Reinforcement Learning as a Parsimonious Alternative to Prediction Cascades: A Case Study on Image Segmentation (AAAI 2024 · similarity 0.83)
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.