# Open-Set Likelihood Maximization for Few-Shot Learning

**Authors:** Malik Boudiaf, Etienne Bennequin, Myriam Tami, Antoine Toubhans, Pablo Piantanida, Céline Hudelot, Ismail Ben Ayed  
**arXiv:** [2301.08390](https://arxiv.org/abs/2301.08390)  
**Date:** 2023-01-20  
**Summary page:** https://www.summarizepaper.com/en/paper/2301.08390v1/

## Key points

- Few-Shot Open-Set Recognition (FSOSR) problem addressed
- Approach called Open-Set Likelihood Optimization (OSLO) introduced
- OSLO extends Maximum Likelihood framework in transductive setting
- Co-optimization of latent scores and parametric models through block-coordinate descent approach
- Interpretability and modularity of OSLO highlighted
- Superior performance of OSLO demonstrated in experiments across five datasets
- Comparison with five state-of-the-art approaches conducted to evaluate effectiveness
- Contributions include introducing OSLO, demonstrating its effectiveness, and evaluating transductive methods for FSOSR

## Summary

, , , , 


In this study, we address the Few-Shot Open-Set Recognition (FSOSR) problem by classifying instances among a limited number of classes with only a few labeled samples while also detecting instances that do not belong to any known class. Our approach, called Open-Set Likelihood Optimization (OSLO), extends the Maximum Likelihood framework in the transductive setting to handle outliers and improve performance. Through a block-coordinate descent approach, we co-optimize latent scores and parametric models iteratively for better results. One key advantage of OSLO is its interpretability and modularity, making it applicable on top of any pre-trained model. Extensive experiments across five datasets demonstrate OSLO's superiority over existing methods in both inlier classification and outlier detection tasks. Furthermore, our study provides valuable insights into the performance of transductive methods for FSOSR through benchmarking against five state-of-the-art approaches. Overall, our contributions include introducing OSLO as a principled extension of Maximum Likelihood for open-set recognition tasks, demonstrating its effectiveness through experiments, and evaluating transductive methods for FSOSR.