Unlocking Transfer Learning for Open-World Few-Shot Recognition

Conference: NeurIPS 2025 (Workshop) arXiv: 2411.09986 Code: N/A Area: Few-Shot Learning / Open-World Recognition Keywords: Few-shot open-set recognition, transfer learning, meta-learning, open world, pseudo open-set samples

TL;DR

A two-stage framework is proposed that combines open-set-aware meta-learning with open-set-free transfer learning, achieving the first successful application of the transfer learning paradigm to few-shot open-set recognition (FSOSR) and reaching SOTA on miniImageNet and tieredImageNet.

Background & Motivation

Few-Shot Open-Set Recognition (FSOSR) presents a dual challenge:

Closed-set classification: Classifying inputs into known categories using only a handful of examples

Open-set detection: Simultaneously identifying inputs that do not belong to any known category

Although transfer learning has become the dominant paradigm for closed-set few-shot learning, its application to open-world scenarios exposes a fundamental limitation: fine-tuning on closed-set samples alone provides no modeling of open-set boundaries. This paper aims to "unlock" the potential of transfer learning for FSOSR.

Method

Overall Architecture

A two-stage pipeline: 1. Stage 1 — Open-Set-Aware Meta-Learning: Trains the model to establish a metric space conducive to open-set detection. 2. Stage 2 — Open-Set-Free Transfer Learning: Fine-tunes the model for specific tasks via transfer learning.

Key Designs

1. Open-Set-Aware Meta-Learning:

   • Introduces simulated open-set samples during the meta-learning stage.
   • Trains the model to not only discriminate among known classes but also recognize unseen inputs.
   • The resulting metric space provides a favorable initialization for subsequent transfer learning.

2. Pseudo Open-Set Sample Generation Strategies:

   • Dataset modification: A subset of training classes is held out to simulate open-set inputs.
   • Generation: Out-of-distribution samples are synthesized in feature space.
   • The two strategies are complementary and can be used jointly.

3. Open-Set-Free Transfer Learning:

   • At test time, fine-tuning is performed using only the closed-set support set.
   • No additional open-set calibration samples are required.
   • Open-set detection capability is maintained through the metric space established in Stage 1.

Loss & Training

\[\mathcal{L} = \mathcal{L}_{\text{CE}} + \lambda \mathcal{L}_{\text{open}}\]

• \(\mathcal{L}_{\text{CE}}\): Cross-entropy loss for closed-set classification
• \(\mathcal{L}_{\text{open}}\): Open-set detection loss (margin-based contrastive loss)

Key Experimental Results

Main Results

Method	miniImageNet AUROC ↑	miniImageNet Acc ↑	tieredImageNet AUROC ↑	tieredImageNet Acc ↑
ProtoNet + threshold	68.5	65.2	70.2	67.8
PEELER	72.3	68.5	73.8	70.5
SnaTCHer	74.1	70.2	75.5	72.3
OpenFSL	75.8	71.5	77.2	73.8
TANE	76.5	72.8	78.1	74.5
Ours	79.2	75.5	80.8	77.2

Training Efficiency Comparison

Method	Training Epochs	Additional Overhead	5-shot AUROC	1-shot AUROC
PEELER	200	+50%	72.3	65.8
SnaTCHer	200	+30%	74.1	67.2
OpenFSL	200	+40%	75.8	68.5
Ours	200	+1.5%	79.2	72.5

Ablation Study

Configuration	AUROC ↑	Closed-set Acc ↑
Full method	79.2	75.5
Meta-learning only (no transfer learning)	75.5	70.2
Transfer learning only (no meta-learning)	71.8	74.8
Without pseudo open-set samples	74.5	73.5
Dataset modification pseudo open-set	77.8	74.8
Generative pseudo open-set	78.5	75.2
Both pseudo open-set strategies	79.2	75.5

Key Findings

1. Transfer learning is indeed viable for FSOSR, provided the pre-training is open-set aware.
2. The additional training overhead is only 1.5%, far lower than competing methods, indicating high efficiency.
3. The two pseudo open-set generation strategies are complementary, with joint use yielding the best results.
4. The metric space established in Stage 1 is foundational to the success of transfer learning.

Highlights & Insights

• Paradigm shift: This work is the first to demonstrate the viability of transfer learning for FSOSR.
• Minimal overhead: SOTA performance is achieved with only a 1.5% training increment.
• Practical strategy: Pseudo open-set sample generation requires no additional data.

Limitations & Future Work

1. As a workshop paper, the experimental scope remains limited.
2. Validation on larger-scale datasets (e.g., full ImageNet) is insufficient.
3. The quality of pseudo open-set samples leaves room for improvement.
4. Integration with pre-trained vision-language models such as CLIP warrants further exploration.

Related Work & Insights

• ProtoNet: A classical metric-learning approach.
• PEELER: A pioneering work on open-set few-shot learning.
• SnaTCHer: Performs open-set detection by "snatching" prototypes.
• CLIP-based FSOSR: Exploits the open-vocabulary capabilities of pre-trained VLMs.

Rating

Dimension	Score (1–5)
Novelty	4
Theoretical Depth	3
Experimental Thoroughness	4
Writing Quality	3
Practical Value	4
Overall Recommendation	3.5

Related Papers

• [NeurIPS 2025] Open-Insect: Benchmarking Open-Set Recognition of Novel Species in Biodiversity Monitoring
• [NeurIPS 2025] Benchmarking Large Language Models for Zero-Shot and Few-Shot Phishing URL Detection
• [NeurIPS 2025] A High-Dimensional Statistical Method for Optimizing Transfer Quantities in Multi-Source Transfer Learning
• [ICCV 2025] Few-Shot Pattern Detection via Template Matching and Regression
• [ICCV 2025] Rethinking Few Shot CLIP Benchmarks: A Critical Analysis in the Inductive Setting