Research map / Label-efficient and robust learning / Recognition and long-tail
Deep Feature Space: A Geometrical Perspective
Ioannis Kansizoglou, Loukas Bampis, Antonios C. Gasteratos
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 |
|---|---|---|---|
| train fusion model on held-out half D_2 instead of data used for extractors (S_a-a -> S_1-2), DNN (ablation) | fusion model · RML + BAUM-1s | mean accuracy (%): 55.64 -> 61.46 | Improved |
| same change with SVM fusion (ablation) | fusion model · RML + BAUM-1s | mean accuracy (%): 59.62 -> 64.05 | Improved |
| same change with extractors pre-trained on AffectNet/IEMOCAP (S'_a-a -> S'_1-2) (ablation) | DNN fusion · RML + BAUM-1s | mean accuracy (%): 65.18 -> 68.98 | Improved |
| add L2 normalisation (L2-Softmax s=20) to Softmax (ablation) | CNN on CIFAR · CIFAR feature space | separability S_R: .990 -> .845 | Hurt |
| lower L2-Softmax scale s from 20 to 0.1 (ablation) | L2-Softmax · CIFAR feature space | separability S_R: .845 -> .588 | Hurt |
| replace Softmax with L2-Softmax s=0.1 (ablation) | CNN on CIFAR · CIFAR feature space | loss-ratio L_R: 2.63 -> 19.52 | Hurt |
| replace Softmax with SphereFace (ablation) | CNN on CIFAR · CIFAR feature space | centrality C_R: .859 -> .949 | Improved |
| replace Softmax with SphereFace (ablation) | CNN on CIFAR · CIFAR feature space | loss-ratio L_R: 2.63 -> 1.24 | Improved |
| replace Softmax with L-Softmax (ablation) | CNN on CIFAR · CIFAR feature space | loss-ratio L_R: 2.63 -> 1.42 | Improved |
| use fusion training strategy S_1-2 vs S_1-1 (DNN) | fusion model · RML + BAUM-1s | mean accuracy (%): 58.53 -> 61.46 | Improved |
| swap DNN fusion for SVM fusion under S_a-a | fusion model · RML + BAUM-1s | mean accuracy (%): 55.64 -> 59.62 | Improved |
| S_a-a vs S_1-2 on speaker s_3 (extractor already well generalised) (ablation) | fusion strategy · RML s_3 DNN | accuracy (%): 70.07 -> 73.05 | Hurt |
Closest papers by meaning
- Deciphering the Feature Representation of Deep Neural Networks for High-Performance AI (TPAMI 2024 · similarity 0.86)
- Provable Multi-Task Representation Learning by Two-Layer ReLU Neural Networks (ICML 2024 · similarity 0.83)
- Understanding Distributed Representations of Concepts in Deep Neural Networks without Supervision (AAAI 2024 · similarity 0.83)
- Improved Generalization of Weight Space Networks via Augmentations (ICML 2024 · similarity 0.83)
- From Layers to States: A State Space Model Perspective to Deep Neural Network Layer Dynamics (ICLR 2025 · similarity 0.83)
- Feature Contamination: Neural Networks Learn Uncorrelated Features and Fail to Generalize (ICML 2024 · similarity 0.82)
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.
