Source-linked AI summary
Adversarial Feature Hallucination Networks for Few-Shot Learning
Kai Li, Yulun Zhang, Kunpeng Li, Yun Fu
TL;DR
Few-shot learning must learn new-class concepts from few labeled samples, while existing augmentation methods do not reliably preserve synthesized-feature discriminability and diversity. AFHN uses conditional Wasserstein feature hallucination with classification and anti-collapse regularizers, and reports superiority across benchmark comparisons.
Problem
Few-shot learning seeks to learn new-class concepts from a few labeled samples, while existing synthesized data may lack discriminability and diversity.
Method
AFHN uses a conditional Wasserstein GAN to synthesize features conditioned on few labeled samples, with classification and anti-collapse regularizers.
Results
AFHN achieves significant improvements over existing state-of-the-art methods on two of three datasets and is comparable on the remaining dataset.
Takeaways & Limitations
AFHN's synthesized features support effective classifiers and improved few-shot recognition results within the evaluated benchmarks.
Abstract
from arXiv · showhide
The recent flourish of deep learning in various tasks is largely accredited to the rich and accessible labeled data. Nonetheless, massive supervision remains a luxury for many real applications, boosting great interest in label-scarce techniques such as few-shot learning (FSL), which aims to learn concept of new classes with a few labeled samples. A natural approach to FSL is data augmentation and many recent works have proved the feasibility by proposing various data synthesis models. However, these models fail to well secure the discriminability and diversity of the synthesized data and thus often produce undesirable results. In this paper, we propose Adversarial Feature Hallucination Networks (AFHN) which is based on conditional Wasserstein Generative Adversarial networks (cWGAN) and hallucinates diverse and discriminative features conditioned on the few labeled samples. Two novel regularizers, i.e., the classification regularizer and the anti-collapse regularizer, are incorporated into AFHN to encourage discriminability and diversity of the synthesized features, respectively. Ablation study verifies the effectiveness of the proposed cWGAN based feature hallucination framework and the proposed regularizers. Comparative results on three common benchmark datasets substantiate the superiority of AFHN to existing data augmentation based FSL approaches and other state-of-the-art ones.
1. Introduction
Few-shot learning addresses classification with scarce labels, and data augmentation seeks to synthesize additional samples. AFHN uses conditional feature hallucination with separate regularizers for synthesized-feature discriminability and diversity.
- Few-shot learning learns concepts of new classes from only a few labeled samples, motivating methods for label-scarce applications.
- Data augmentation methods synthesize or transform images and features by transferring intra-class cross-sample relationships from seen to unseen classes.The augmented samples are intended to diversify intra-class variance and sharpen classification boundaries.
- Existing augmentation methods struggle to guarantee both discriminability and diversity: arbitrary mappings can damage class information, while task constraints can cause mode collapse.
- AFHN combines a conditional Wasserstein GAN feature-synthesis framework with a classification regularizer that encourages same-class similarity and different-class dissimilarity.The seed labeled features provide the conditional context for hallucinating new features.
- AFHN adds an anti-collapse regularizer that penalizes likely mode collapse by relating synthesized-feature dissimilarity to latent-noise dissimilarity.The constraint encourages the generator to explore minor distribution modes.
- AFHN reports state-of-the-art performance on three common benchmark datasets.
2. Related Work
Few-shot learning research mainly improves feature discriminability, model adaptability, or data availability through augmentation. AFHN belongs to the augmentation category and differs from related GAN and feature-hallucination methods in architecture, regularization, and classifier use.
- Feature discriminability: Few-shot learning methods enhance feature discriminability using metric learning and related adaptive similarity mechanisms.
- Model flexibility: Meta-learning methods learn algorithms that produce models readily updated for new tasks from limited task information.
- Data augmentation: Data augmentation methods address label scarcity through external-data transfer or transformations of labeled images and feature representations.
- AFHN: AFHN synthesizes fake features conditioned on labeled-sample features and uses a Wasserstein GAN rather than the conventional GAN framework.
- AFHN: Unlike prior GAN-based models, AFHN includes a classification regularizer, a novel anti-collapse regularizer, and a standard Softmax classifier trained on synthesized features.
- AFHN: Compared with a prior feature-hallucination method, AFHN uses an adversarial generative model, trains a standard Softmax classifier, and explicitly targets feature diversity.
3. Algorithm
AFHN builds a conditional WGAN-based few-shot feature hallucination pipeline that generates class-conditioned features from support-set representations and trains a classifier with them. Classification and anti-collapse regularizers target discriminability and diversity, while generated features augment novel-class support sets at test time.
- Conditional WGAN framework: AFHN extracts support-set features, conditions generator G on class prototypes, and combines them with two noise variables to synthesize fake features.For multiple samples per class, their feature vectors are averaged into a prototype before synthesis.
- Conditional WGAN framework: WGAN training makes discriminator D distinguish real support features from generated features, producing the adversarial loss L_GAN.The WGAN formulation uses Wasserstein distance and gradient-norm penalization.
- Regularization: The classification regularizer encourages synthesized features to match their conditioned class while remaining dissimilar from features of other classes.The classifier uses cosine similarity between query features and synthesized features.
- Regularization: The anti-collapse regularizer penalizes similar latent codes producing collapsed features, amplifying dissimilarity and encouraging exploration of minor feature-space modes.The final objective combines the regularized terms with hyperparameters α and β.
- Classification with synthesized samples: At test time, AFHN augments each novel-class labeled support set with generated features, then trains a Softmax classifier to classify query samples.The test task uses classes without overlap with the training classes.
4. Experiments
AFHN is evaluated on three few-shot benchmarks, with ablations testing its conditional WGAN framework and two regularizers. Results show improved classification, stronger feature structure, and generally superior performance to existing methods.
- Experimental Setup: AFHN is evaluated on Mini-ImageNet, CUB, and CIFAR100 using standard 5-way 1-shot and 5-way 5-shot tasks.Evaluation averages accuracy over 600 sampled tasks with 95% confidence intervals.
- Ablation Study: CR raises Mini-ImageNet 5-way 1-shot accuracy from 52.73 to 55.65 without cWGAN and from 57.58 to 60.56 with cWGAN.The ablation attributes these gains to improved discriminability of synthesized features.
- Ablation Study: AR adds about 2% performance gain by explicitly penalizing likely mode collapse, supporting its role in diversifying synthesized features.The regularizer targets diversity rather than classification structure.
- Comparative Results: On Mini-ImageNet, AFHN exceeds Δ-encoder by more than 8% in 5-shot and Dual TriNet by more than 3% in 1-shot accuracy.It also improves over MetaGAN by about 10% in both settings.
- Comparative Results: Across the three datasets, AFHN significantly improves over existing state-of-the-art methods on two datasets and is comparable on the remaining dataset.It significantly improves over the no-augmentation baseline on all three datasets.
- Further Analysis: With CR, synthesized features form clearer class clusters; adding AR preserves clustering while expanding cluster regions and introducing more variation.Using only cWGAN produces mixed class embeddings, whereas AR increases feature diversity.
5. Conclusions
AFHN combines cWGAN-based feature synthesis with classification and anti-collapse regularizers to produce discriminative and diverse features for few-shot learning. Ablation and comparative results support the effectiveness and superiority of the approach.
- AFHN synthesizes fake features through a cWGAN framework conditioned on features from a few labeled samples.
- The classification regularizer promotes discriminability by encouraging similarity within classes and dissimilarity across classes.
- The anti-collapse regularizer enhances feature diversity by penalizing likely mode-collapse cases.
- Ablation studies verify the effectiveness of the cWGAN feature-synthesis framework and both regularizers.
- Comparative results verify AFHN’s superiority over existing data-augmentation-based FSL approaches and other state-of-the-art methods.