Scholarlink

<p>A&#xa0;method for solving the concept-based learning (CBL) problem is proposed. An idea behind the method is to divide each concept-annotated image into patches, transform them into embeddings using an autoencoder, and cluster the embeddings assuming that each cluster will primarily contain embeddings of patches representing certain concepts. To infer concepts for a&#xa0;new image, the method employs frequentist inference by computing prior and posterior probabilities of concepts based on frequencies of patches from images associated with concept values. The proposed method is named Frequentist Inference CBL (FI-CBL). It enables the incorporation of expert rules in a&#xa0;form of logical functions into the inference process by updating probabilities of concepts to ensure consistency with these rules. FI-CBL offers a&#xa0;clear frequency-based interpretation. Numerical experiments show that FI-CBL outperforms the concept bottleneck model, particularly when the amount of training data is limited. The code of proposed algorithms is publicly available.</p>

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

FI-CBL: A Probabilistic Method for Concept-Based Learning with Expert Rules

  • 期刊论文

  • 发表日期：

    2025年4月1日

查看全文
  • Lev V. Utkin,
  • Andrei V. Konstantinov,
  • Arslan A. Kaygermazov,
  • Stanislav R. Kirpichenko,
  • Natalya M. Verbova,
  • Vladimir A. Muliukha

摘要

A method for solving the concept-based learning (CBL) problem is proposed. An idea behind the method is to divide each concept-annotated image into patches, transform them into embeddings using an autoencoder, and cluster the embeddings assuming that each cluster will primarily contain embeddings of patches representing certain concepts. To infer concepts for a new image, the method employs frequentist inference by computing prior and posterior probabilities of concepts based on frequencies of patches from images associated with concept values. The proposed method is named Frequentist Inference CBL (FI-CBL). It enables the incorporation of expert rules in a form of logical functions into the inference process by updating probabilities of concepts to ensure consistency with these rules. FI-CBL offers a clear frequency-based interpretation. Numerical experiments show that FI-CBL outperforms the concept bottleneck model, particularly when the amount of training data is limited. The code of proposed algorithms is publicly available.