Research map / Architectures and efficiency / CNN design
Feedback-Prop: Convolutional Neural Network Inference Under Partial Evidence
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 layer-wise feedback-prop with known coarse-scene labels | AlexNet scene classifier · SUN397 | MC Acc: 52.83$\pm$0.24 -> 54.93$\pm$0.42 | Helped |
| Use residual feedback-prop with known coarse-scene labels | AlexNet scene classifier · SUN397 | mAP: 56.17$\pm$0.21 -> \bf{58.70$\pm$0.26} | Helped |
| Provide a larger portion of surrounding news text as known evidence | ResNet-50 news-image concept prediction · News images test | mAP: 19.92 -> 22.57 | Helped |
| Use object annotations as evidence with LF at Conv10 | ResNet-50 captioning model · COCO captioning test | BLEU-4: 28.65 -> 29.78 | Helped |
| Use object annotations as evidence with RF at Conv10 and Conv40 | ResNet-50 captioning model · COCO captioning test | CIDEr: 0.9466 -> 0.9922 | Helped |
| Use 50 known labels as partial evidence | VGG-16 multi-label model with LF at Conv13 · COCO multi-label test | mAP: 27.09 -> 27.41 | Helped |
| Use 500 known labels as partial evidence | VGG-16 multi-label model · COCO multi-label test | mAP: 27.09 -> 29.76 | Helped |
| Use middle convolutional layers rather than layers near the input or output as pivots (ablation) | Feedback-prop multi-label models · COCO multi-label task | mAP | Mixed |
| Use Conv13 instead of no feedback-prop (ablation) | ResNet-18 with LF · COCO multi-label task | mAP: 24.08 -> 26.80 | Helped |
| Use Conv13 instead of no feedback-prop (ablation) | ResNet-18 with RF · COCO multi-label task | mAP: 24.08 -> 27.26 | Helped |
| Use Conv13 rather than Conv7 as the pivot (ablation) | VGG-16 with LF · COCO multi-label task | mAP: 29.82 -> 27.59 | Hurt |
| Use multiple pivots, Conv10 and Conv40, rather than a single pivot (ablation) | ResNet-50 with RF · COCO captioning test | CIDEr: 0.9466 -> 0.9922 | Helped |
Closest papers by meaning
- Neural Networks with Recurrent Generative Feedback (NeurIPS 2020 · similarity 0.87)
- Involution: Inverting the Inherence of Convolution for Visual Recognition (CVPR 2021 · similarity 0.87)
- Approximation and Learning with Deep Convolutional Models: a Kernel Perspective (ICLR 2022 · similarity 0.86)
- Task-Driven Convolutional Recurrent Models of the Visual System (NeurIPS 2018 · similarity 0.86)
- CNN Filter DB: An Empirical Investigation of Trained Convolutional Filters (CVPR 2022 · similarity 0.86)
- PICNN: A Pathway towards Interpretable Convolutional Neural Networks (AAAI 2024 · 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.