# Test-Time Discovery via Hashing Memory

**Authors:** Fan Lyu, Tianle Liu, Zhang Zhang, Fuyuan Hu, Liang Wang  
**arXiv:** [2503.10699](https://arxiv.org/abs/2503.10699)  
**Date:** 2025-03-12  
**Summary page:** https://www.summarizepaper.com/en/paper/2503.10699v1/

## Key points

- Test-Time Discovery (TTD) introduced as a novel task to address class shifts during testing
- Challenge in TTD: distinguishing newly discovered classes from known ones
- Proposed solution: training-free, hash-based memory mechanism for fine-grained comparisons with past samples
- Leveraging characteristics of unknown classes through hash representation based on feature scale and directions
- Utilization of Locality-Sensitive Hashing (LSH) for efficient grouping of similar samples
- Collaborative classification strategy combining prototype classifier for known classes with LSH-based classifier for novel ones
- Incorporation of self-correction mechanism refining memory labels through hash-based neighbor retrieval for stable and accurate class assignments
- Experimental results show effective discovery of novel categories while maintaining performance on known classes
- References related works such as gradient episodic memory and variational continual test-time adaptation
- Illustrations in the paper depict boundary between prototype and LSH classifiers, trends of TA and CA across TTD process
- Proposed method demonstrates effective hashing strategies for test-time discovery and classification tasks

## Summary

In their paper titled "Test-Time Discovery via Hashing Memory," Fan Lyu, Tianle Liu, Zhang Zhang, Fuyuan Hu, and Liang Wang introduce Test-Time Discovery (TTD) as a novel task aimed at addressing class shifts during testing. The task requires models to identify emerging categories while preserving previously learned ones. A key challenge in TTD is distinguishing newly discovered classes from those already identified. To tackle this challenge, the authors propose a training-free, hash-based memory mechanism that enhances class discovery through fine-grained comparisons with past test samples. Their approach leverages the characteristics of unknown classes by introducing hash representation based on feature scale and directions. They utilize Locality-Sensitive Hashing (LSH) for efficient grouping of similar samples, enabling test samples to be easily and quickly compared with relevant past instances. Additionally, the authors design a collaborative classification strategy that combines a prototype classifier for known classes with an LSH-based classifier for novel ones. To enhance reliability, they incorporate a self-correction mechanism that refines memory labels through hash-based neighbor retrieval, ensuring more stable and accurate class assignments. Experimental results demonstrate that their method achieves good discovery of novel categories while maintaining performance on known classes, establishing a new paradigm in model testing. The paper also references related works such as gradient episodic memory for continual learning and variational continual test-time adaptation. Figures in the paper illustrate the boundary between prototype and LSH classifiers as well as trends of TA (test-time adaptation) and CA (classification accuracy) across the entire TTD process. Overall, the proposed method showcases effective hashing strategies for test-time discovery and classification tasks. The detailed experimental results provide insights into the performance of the approach and its potential applications in real-world scenarios.