Fig. 4: Comprehensive training dynamics and multi-metric analysis during MAGPIE pretraining. | npj Digital Medicine

Fig. 4: Comprehensive training dynamics and multi-metric analysis during MAGPIE pretraining.

From: Masked autoencoding, generalizable pretraining, and integrated experts for enhanced glioma segmentation

Fig. 4: Comprehensive training dynamics and multi-metric analysis during MAGPIE pretraining.

Top-left: Multi-metric evolution heatmap showing the progression of various training metrics (train/validation loss, BLEU, ROUGE-L, expert utilization, attention, learning rate, gradient norm) across 50 epochs. Top-right: Training and validation loss curves demonstrating convergence behavior. Middle-right: Quality metrics (BLEU and ROUGE-L scores) evolution over epochs. Bottom-left: Expert utilization heatmap showing dynamic activation patterns of 8 experts (E1-E8) throughout training. Bottom-middle: Learning rate schedule with step-decay strategy. Bottom-right: Gradient stability showing gradient norm evolution with clipping threshold.

Back to article page