From Knowledge Transmission to Knowledge Construction: A Step towards Human-Like Active Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence in Education
• Educational Technology

Background:

• Traditional machine learning uses guess-and-check, which is inefficient for large datasets.
• Active learning enhances performance by selecting crucial data, reducing annotation and analysis costs, especially for deep learning.
• Current machine active learning lacks tools for knowledge construction, unlike human-like active learning.

Purpose of the Study:

• To address the gap in machine active learning for educational settings by proposing a novel data selection strategy.
• To introduce a method for measuring data information capacity using the four-parameter logistic item response theory (4PL IRT).
• To compare the proposed strategy against existing active learning methods like Least Confidence and Entropy Sampling.

Main Methods:

• Developed a new active learning strategy based on the information function from 4PL IRT.
• Quantified data information capacity to guide the selection of informative samples.
• Conducted computational experiments to evaluate the strategy's effectiveness.

Main Results:

• The proposed Information Capacity strategy demonstrated comparable behavior to Least Confidence and Entropy Sampling.
• The strategy offers a more flexible framework for creating transparent knowledge models in deep learning applications.
• Computational experiments validated the efficacy of the 4PL IRT-based approach in active learning.

Conclusions:

• The Information Capacity strategy provides a robust and flexible method for active learning in educational deep learning.
• This approach enhances transparency in knowledge models, facilitating better understanding and development.
• The study highlights the potential of integrating item response theory with machine learning for educational advancements.