Which Samples Should Be Learned First: Easy or Hard?

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Unequal treatment of training samples is common in machine learning.
• Existing weighting schemes often adopt either an easy-first or hard-first approach.
• The optimal sample prioritization strategy for new tasks remains an open question.

Purpose of the Study:

• To investigate whether easy or hard samples should be prioritized in new learning tasks.
• To propose a flexible weighting scheme that adapts to varying data difficulty distributions.
• To provide a comprehensive answer to the 'easy-or-hard' sample prioritization dilemma.

Main Methods:

• Theoretical analysis using a general objective function to derive optimal weights.
• Identification of four distinct sample prioritization modes: easy-first, hard-first, medium-first, and two-ends-first.
• Development and experimental validation of a flexible weighting scheme (FlexW).

Main Results:

• The optimal sample prioritization mode depends on the training set's difficulty distribution.
• Novel prioritization modes (medium-first, two-ends-first) were identified alongside traditional ones.
• FlexW demonstrated effectiveness across various learning scenarios and outperformed existing methods.

Conclusions:

• The choice between easy-first and hard-first learning is task-dependent and influenced by data characteristics.
• The proposed FlexW scheme offers adaptability and optimal performance by dynamically selecting the best prioritization mode.
• This research provides valuable insights for optimizing machine learning training strategies.