# Glocal Energy-based Learning for Few-Shot Open-Set Recognition

**Authors:** Haoyu Wang, Guansong Pang, Peng Wang, Lei Zhang, Wei Wei, Yanning Zhang  
**arXiv:** [2304.11855](https://arxiv.org/abs/2304.11855)  
**Date:** 2023-04-24  
**Summary page:** https://www.summarizepaper.com/en/paper/2304.11855v1/

## Key points

- The paper addresses the challenging task of few-shot open-set recognition (FSOR)
- FSOR involves categorizing a sample into one of the pre-defined, closed-set classes based on limited examples while also being able to reject samples from unknown classes
- The authors propose a novel energy-based hybrid model consisting of a classification branch and an energy branch
- The classification branch learns a metric to classify samples into closed-set classes
- The energy branch estimates the open-set probability, enabling holistic detection of open-set samples
- The model utilizes class-wise features to learn a global energy score and pixel-wise features to learn a local energy score
- High energy scores are assigned to samples that deviate from few-shot examples in either class-wise or pixel-wise features, and low energy scores otherwise
- The model outperforms existing approaches on three standard FSOR datasets
- It accurately categorizes samples into closed set classes while effectively rejecting unknown samples
- The approach has potential for practical applications in various domains.

## Summary

The paper titled "Glocal Energy-based Learning for Few-Shot Open-Set Recognition" addresses the challenging task of few-shot open-set recognition (FSOR), which has significant practical value. FSOR involves categorizing a sample into one of the pre-defined, closed-set classes based on limited examples while also being able to reject samples from unknown classes. To tackle this task, the authors propose a novel energy-based hybrid model. The proposed model consists of two branches: a classification branch and an energy branch. The classification branch learns a metric to classify a sample into one of the closed-set classes, while the energy branch explicitly estimates the open-set probability. This approach enables holistic detection of open-set samples by leveraging both class-wise and pixel-wise features. To learn a glocal energy-based score, the model utilizes class-wise features to learn a global energy score and pixel wise features to learn a local energy score. The model is designed to assign high energy scores to samples that deviate from the few shot examples in either class wise or pixel wise features, and low energy scores otherwise. The authors evaluate their model on three standard FSOR datasets and demonstrate its superior performance compared to existing approaches. The results highlight the effectiveness of their proposed energy based hybrid model in accurately categorizing samples into closed set classes while effectively rejecting unknown samples. In conclusion, this paper presents an innovative approach for few shot open set recognition using an energy based hybrid model. By leveraging both class wise and pixel wise features, their model achieves robust detection of open set samples while maintaining high accuracy in categorizing closed set samples. The experimental results validate the effectiveness of their approach and its potential for practical applications in various domains.