Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
GP-Tree: A Gaussian Process Classifier for Few-Shot Incremental Learning
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 Gibbs sampling GP-Tree with semantic (prototype-clustered) tree instead of random balanced tree (ablation) | GP-Tree on CUB pre-trained features · CUB-200-2011, 4-60 classes | test accuracy: GP-Tree 69.92 vs Rnd 58.49 at 60 classes | Helped |
| Replace stick-breaking sequential tree with semantic hierarchical tree (GP-Tree) | GP classification with Polya-Gamma · CUB-200-2011, 60 classes | test accuracy: 69.92 vs 60.69 (SBM-GP) | Helped |
| GP-Tree tree-based hierarchical classification vs LSM/OVE at few classes | GP-Tree · CUB-200-2011, <=30 classes | test accuracy | No clear effect |
| Add DKL with GP-Tree (VI, inducing points) instead of SV-DKL / GPDNN | ResNet-18 on CIFAR-10 and CIFAR-100 · CIFAR-10 / CIFAR-100 | test accuracy: 93.25 vs 92.73 (CIFAR-10); 72.11 vs 70.61 (CIFAR-100) | Helped |
| Use GP-Tree with Gibbs-sampled novel-session nodes in few-shot class-incremental learning | GP-Tree FSCIL · CUB-200-2011 FSCIL, sessions 4-11 | test accuracy: session 11: 44.48 vs 39.30 (PODNet); session 4: 59.61 vs 49.00 | Helped |
| Use GP-Tree instead of PODNet in early FSCIL sessions | GP-Tree FSCIL · CUB-200-2011 FSCIL, session 1-3 | test accuracy: session 1: 73.73 vs 75.93 (PODNet); session 2: 68.24 vs 70.29 | Hurt |
| Build tree per session from current-session classes only instead of all novel sessions' samples (ablation) | GP-Tree FSCIL tree construction · CUB-200-2011 FSCIL, sessions 2-11 | test accuracy: session 11: 43.33 vs 44.48 | Hurt |
| Rebuild entire tree each session with base-class inducing inputs instead of retaining base tree (ablation) | GP-Tree FSCIL tree construction · CUB-200-2011 FSCIL, sessions 2-11 | test accuracy: session 11: 43.20 vs 44.48 | Hurt |
| Use Gibbs sampling instead of variational inference for GP-Tree inference (ablation) | GP-Tree on CUB pre-trained features · CUB-200-2011, 4-60 classes | test accuracy: large performance gap in favor of Gibbs, widening with more classes | Helped |
| Use RBF/Matern kernel instead of linear kernel for novel sessions (ablation) | GP-Tree FSCIL · CUB-200-2011 FSCIL, novel sessions | test accuracy | Helped |
| Reduce representative samples per class from 5 to 1-4 (ablation) | GP-Tree FSCIL · CUB-200-2011 FSCIL | test accuracy: slight degradation, more severe in later sessions | Hurt |
Closest papers by meaning
- Scalable Multi-Class Gaussian Process Classification using Expectation Propagation (ICML 2017 · similarity 0.87)
- Efficient Transformed Gaussian Processes for Non-Stationary Dependent Multi-class Classification (ICML 2023 · similarity 0.86)
- Log-Linear-Time Gaussian Processes Using Binary Tree Kernels (NeurIPS 2022 · similarity 0.86)
- Efficient Gaussian Process Classification Using Pólya-Gamma Data Augmentation (AAAI 2019 · similarity 0.86)
- Learning Gaussian Processes by Minimizing PAC-Bayesian Generalization Bounds (NeurIPS 2018 · similarity 0.86)
- Gaussian Process Neural Additive Models (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.