Research map / Label-efficient and robust learning / Recognition and long-tail
Introspective Classification with Convolutional Nets
Long Jin, Justin Lazarow, Zhuowen Tu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 iteratively synthesized pseudo-negatives rather than using only the CNN training set | CNN · MNIST, one-vs-all | test error (%): 0.87 -> 0.78 | Improved |
| replace meaningful synthesized pseudo-negatives with random noise (ablation) | ICN · MNIST, one-vs-all | test error (%): 0.78 -> 0.89 | Hurt |
| add iteratively synthesized pseudo-negatives | ResNet-32 · MNIST, softmax | test error (%): 0.50 -> 0.47 | Improved |
| add iteratively synthesized pseudo-negatives | ResNet-32 · CIFAR-10, no data augmentation, one-vs-all | test error (%): 13.44 -> 12.94 | Improved |
| add iteratively synthesized pseudo-negatives | ResNet-32 · CIFAR-10, no data augmentation, softmax | test error (%): 12.38 -> 11.46 | Improved |
| add iteratively synthesized pseudo-negatives | ResNet-32 · CIFAR-10, data augmentation, softmax | test error (%): 7.06 -> 6.70 | Improved |
| use random noise instead of classifier-synthesized pseudo-negatives (ablation) | ICN with ResNet-32 · CIFAR-10, no data augmentation, softmax | test error (%): 11.46 -> 11.94 | Hurt |
| add iteratively synthesized pseudo-negatives | ResNet-32 · SVHN, softmax | test error (%): 2.01 -> 1.95 | Improved |
| choose between SGD and SGD Langevin sampling, or the listed stopping options (ablation) | ICN · MNIST, classification test error | test error (%): 0.78 -> 0.80; 0.81; 0.72; 0.74; 0.75; 0.73 | No clear effect |
| use label smoothing | ResNet-32 · CIFAR-10, no data augmentation, softmax | test error (%): 12.38 -> 12.65 | Hurt |
Closest papers by meaning
- Introspective Neural Networks for Generative Modeling (ICCV 2017 · similarity 0.90)
- Introspective Learning : A Two-Stage approach for Inference in Neural Networks (NeurIPS 2022 · similarity 0.86)
- i-RevNet: Deep Invertible Networks (ICLR 2018 · similarity 0.85)
- Invertibility of Convolutional Generative Networks from Partial Measurements (NeurIPS 2018 · similarity 0.85)
- PICNN: A Pathway towards Interpretable Convolutional Neural Networks (AAAI 2024 · similarity 0.84)
- Convolutional Channel-Wise Competitive Learning for the Forward-Forward Algorithm (AAAI 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.
