New Framework Improves Data Efficiency in Curriculum Learning

This new research introduces a framework designed to enhance curriculum learning by separating how training samples are scored for difficulty from how these samples are introduced into the training process. This disentanglement allows for a clearer understanding of which components contribute to performance gains. The proposed "confusion-aware" difficulty score considers both the model's confidence in correct classifications and its distribution of probabilities across incorrect classes. While this score produces intuitive difficulty rankings, the study found that simply improving the scoring function doesn't necessarily boost accuracy with full datasets. However, when applied in a curriculum ordering, this confusion-aware approach significantly improves data efficiency, especially in scenarios with limited data. It achieved up to an 8.7% point increase in performance compared to random data ordering when using only 20% of the data, highlighting its potential for more efficient model training.