Source-linked AI summary
Generating Classification Weights with GNN Denoising Autoencoders for Few-Shot Learning
Spyros Gidaris, Nikos Komodakis
TL;DR
The paper addresses few-shot adaptation of a recognition model to classify both novel and base classes by generating classification weights from limited examples. It uses a denoising autoencoder implemented as a GNN and reports state-of-the-art results across ImageNet-FS, MiniImageNet, and tiered-MiniImageNet.
Problem
Few-shot learning requires adapting a model to novel classes from limited examples while retaining unified recognition of previously learned base classes.
Method
A Denoising Autoencoder generates and iteratively refines classification weights, while a GNN models co-dependencies across the task’s class-weight set.
Results
The models achieve state-of-the-art results on ImageNet-FS, MiniImageNet, and tiered-MiniImageNet; wDAE-GNN improves prior 1-shot novel-class accuracy on ImageNet-FS by around 1.8 accuracy points.
Takeaways & Limitations
Noise regularization and graph-based modeling support classification-weight generation that exploits structure across base and novel classes.
Takeaways & Limitations
The approach assumes an initial recognition network has already been trained to recognize a set of base classes.
Abstract
from arXiv · showhide
Given an initial recognition model already trained on a set of base classes, the goal of this work is to develop a meta-model for few-shot learning. The meta-model, given as input some novel classes with few training examples per class, must properly adapt the existing recognition model into a new model that can correctly classify in a unified way both the novel and the base classes. To accomplish this goal it must learn to output the appropriate classification weight vectors for those two types of classes. To build our meta-model we make use of two main innovations: we propose the use of a Denoising Autoencoder network (DAE) that (during training) takes as input a set of classification weights corrupted with Gaussian noise and learns to reconstruct the target-discriminative classification weights. In this case, the injected noise on the classification weights serves the role of regularizing the weight generating meta-model. Furthermore, in order to capture the co-dependencies between different classes in a given task instance of our meta-model, we propose to implement the DAE model as a Graph Neural Network (GNN). In order to verify the efficacy of our approach, we extensively evaluate it on ImageNet based few-shot benchmarks and we report strong results that surpass prior approaches. The code and models of our paper will be published on: https://github.com/gidariss/wDAE_GNN_FewShot
1. Introduction
The paper targets few-shot adaptation of a recognition model while preserving performance on base classes, using a denoising autoencoder and graph neural network to generate discriminative classification weights. It reports state-of-the-art results across several ImageNet-based benchmarks.
- Research problem: Few-shot object recognition seeks to recognize classes from only a few examples by training meta-learning systems across related tasks.These methods exploit knowledge acquired from other tasks to compensate for scarce training data.
- Research problem: The proposed meta-model adapts a recognition network trained on base classes to novel classes while maintaining recognition performance on the base classes.The target setting includes very few examples per novel class, such as one or five.
- Denoising weight generation: During training, a Denoising Autoencoder reconstructs target-discriminative classification weights from inputs corrupted with additive Gaussian noise.The injected noise regularizes the weight-generation meta-model and helps reduce overfitting.
- Graph-based modeling: The DAE is implemented as a Graph Neural Network to exploit co-dependencies among the entire set of base and novel class weight vectors.This uses relationships among semantically or visually similar classes rather than reconstructing each class independently.
- Contributions and evaluation: The contributions include applying DAEs to few-shot learning, using GNNs for classification-weight generation, and detailed evaluation on ImageNet-FS and MiniImageNet.The paper reports state-of-the-art results on ImageNet-FS, MiniImageNet, and tiered-MiniImageNet.
2. Methodology
The method adapts a base-class recognition model to jointly classify base and novel classes by generating and refining classification weights from few-shot data. It combines a denoising autoencoder with a GNN to regularize generation and model dependencies among classes.
- Recognition model: The classifier uses L2-normalized features and class weight vectors, with scores computed through cosine similarity.For an image feature z, class scores are computed as z^T w_i.
- Weight generation: The meta-model generates classification weights for both base and novel classes so the adapted classifier recognizes them jointly.It uses the existing base-class weights and few examples from each novel class.
- Denoising autoencoder: The DAE receives noisy initial weight estimates and learns to reconstruct target-discriminative weights, using noise to regularize the generation model.Its reconstruction can iteratively refine initial estimates toward more probable configurations conditioned on the available training data.
- Inference: Initial weights retain the learned base-class vectors and average feature vectors from the few training examples for each novel class.The resulting estimates are refined by repeatedly applying the classification-weight DAE.
- Graph-based denoising: A GNN implements the DAE so reconstruction can use inter-class relationships while respecting the unordered structure of class sets.Training uses episodes in which sampled base classes are treated as fake novel classes, with separate few-shot and validation examples.
3. Experimental Evaluation
The evaluation tests the method on three few-shot benchmarks using novel-class and, for some settings, unified novel-and-base-class accuracy. Across these experiments, the proposed models outperform prior methods, while ablations show benefits from denoising and graph-based class dependencies.
- Datasets and Evaluation Metrics: The evaluation covers ImageNet-FS, MiniImageNet, and tiered-MiniImageNet with benchmark-specific novel-class and unified-class recognition tasks.ImageNet-FS reports 311-way novel-class and 507-way all-class recognition; MiniImageNet and tiered-MiniImageNet use 5-way episodes.
- Datasets and Evaluation Metrics: The experiments use multiple K-shot settings, averaging accuracy over 100 ImageNet-FS episodes and 20,000 episodes for the other benchmarks.ImageNet-FS evaluates K = 1, 2, 5, 10, and 20; MiniImageNet and tiered-MiniImageNet evaluate K = 1 and 5.
- Comparison with Prior Work: The proposed models surpass prior methods across MiniImageNet and tiered-MiniImageNet and usually outperform them on ImageNet-FS, especially for K ≤5.On ImageNet-FS, wDAE-GNN improves the previous state of the art by around 1.8 accuracy points in the 1-shot novel-class setting.
- Ablation Study of DAE Framework: Ablations examine input noise, noisy targets, classification and reconstruction losses, and refinement of initial classification-weight estimates.The study compares variants that remove noise, alter the input targets, remove either loss term, or omit DAE refinement.
- Impact of GNN Architecture: The GNN architecture yields a small but consistent improvement over the MLP architecture, around 0.40 percentage points in the 1-shot case.The improvement is statistically significant in almost all cases according to the confidence intervals in Tables 1 and 2.
4. Conclusion
The paper proposes a DAE-based meta-model, implemented with a GNN, to generate and refine classification weights for novel classes from few examples. Experiments report significance for the DAE formulation, consistent GNN gains, and performance surpassing prior methods on the explored datasets.
- Conclusion: The meta-model generates classification weight vectors for novel classes from only a few training examples.It is associated with a recognition model trained on base classes and aims to adapt that model while retaining base-class performance.
- Conclusion: Gaussian noise regularizes DAE training, and the trained model refines initial classification weights to make them more discriminative for the task.The DAE is used during testing to refine an initial set of classification-weight estimates.
- Conclusion: The GNN implementation models the structure and dependencies of the entire set of classification weights reconstructed in each meta-learning episode.This design lets the meta-model use the structure of the complete weight set rather than treating classes independently.
- Conclusion: Experiments reveal the significance of the DAE formulation and consistent few-shot accuracy improvements from the GNN architecture.The conclusion reports these findings from experiments on ImageNet-FS and MiniImageNet.
- Conclusion: The model surpasses prior methods on all explored datasets.The conclusion states this overall result after the reported ImageNet-FS and MiniImageNet analyses.