HYBRID GRANULARITY DISTRIBUTION ESTIMATION FOR FEW-SHOT LEARNING: STATISTICS TRANSFER FROM CATEGORIES AND INSTANCES
Primary Area: representation learning for computer vision, audio, language, and other modalities
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: few shot learning, representation learning, computer vision
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Distribution estimation (DE) is one of the effective strategies for few-shot learning (FSL).
It involves sampling additional training data for novel categories by estimating their distributions employing transferred statistics (*i*.*e*., mean and variance) from similar base categories.
This strategy enhances data diversity for novel categories and leads to effective performance improvement.
However, we argue that relying solely on coarse-grained estimation at category-level fails to generate representative samples due to the discrepancy between the base categories and the novel categories.
To pursue representativeness while maintaining the diversity of the generated samples, we propose **H**ybrid **G**ranularity **D**istribution **E**stimation (HGDE), which estimates distributions at both coarse-grained category and fine-grained instance levels.
In HGDE, apart from coarse-grained category statistics, we incorporate external fine-grained instance statistics derived from nearest base samples to provide a representative description of novel categories. Then we fuse the statistics from different granularity through a linear interpolation to finally characterize the novel categories. Empirical studies conducted on four FSL benchmarks demonstrate the effectiveness of HGDE in improving the recognition accuracy of novel categories.
Furthermore, HGDE can be applied to enhance the classification performance in other FSL methods. The code is available at:
[https://anonymous.4open.science/r/HGDE-2026}](https://anonymous.4open.science/r/HGDE-2026)
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 2165
Loading