Scholarlink

This work presents an approach to generate Sparse Distributed Representations (SDRs) for word representations using singular value decomposition (SVD). SDRs provide a sparse, biologically inspired alternative to traditional dense word embeddings such as Word2Vec and GloVe, with advantages in terms of robustness to noise and reduced memory requirements. By applying SVD to the co-occurrence matrix of a large corpus, latent semantic structures are extracted and converted into sparse vectors. Experiments show that the generated SDRs can partially represent semantic relationships well, but reach their limits with complex or ambiguous terms.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Efficient Generation of Sparse Distributed Representations (SDRs) with Singular Value Decomposition (SVD)

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Martin Drebinger

摘要

This work presents an approach to generate Sparse Distributed Representations (SDRs) for word representations using singular value decomposition (SVD). SDRs provide a sparse, biologically inspired alternative to traditional dense word embeddings such as Word2Vec and GloVe, with advantages in terms of robustness to noise and reduced memory requirements. By applying SVD to the co-occurrence matrix of a large corpus, latent semantic structures are extracted and converted into sparse vectors. Experiments show that the generated SDRs can partially represent semantic relationships well, but reach their limits with complex or ambiguous terms.