A Fresh Look at Generalized Category Discovery through Non-negative Matrix Factorization

  • 2024-10-30 02:34:11
  • Zhong Ji, Shuo Yang, Jingren Liu, Yanwei Pang, Jungong Han
  • 0

Abstract

Generalized Category Discovery (GCD) aims to classify both base and novelimages using labeled base data. However, current approaches inadequatelyaddress the intrinsic optimization of the co-occurrence matrix $\bar{A}$ basedon cosine similarity, failing to achieve zero base-novel regions and adequatesparsity in base and novel domains. To address these deficiencies, we propose aNon-Negative Generalized Category Discovery (NN-GCD) framework. It employsSymmetric Non-negative Matrix Factorization (SNMF) as a mathematical medium toprove the equivalence of optimal K-means with optimal SNMF, and the equivalenceof SNMF solver with non-negative contrastive learning (NCL) optimization.Utilizing these theoretical equivalences, it reframes the optimization of$\bar{A}$ and K-means clustering as an NCL optimization problem. Moreover, tosatisfy the non-negative constraints and make a GCD model converge to anear-optimal region, we propose a GELU activation function and an NMF NCE loss.To transition $\bar{A}$ from a suboptimal state to the desired $\bar{A}^*$, weintroduce a hybrid sparse regularization approach to impose sparsityconstraints. Experimental results show NN-GCD outperforms state-of-the-artmethods on GCD benchmarks, achieving an average accuracy of 66.1\% on theSemantic Shift Benchmark, surpassing prior counterparts by 4.7\%.

 

Quick Read (beta)

loading the full paper ...