Normalized total loss on the Facebook dataset (DCOR with and without RLC).
<p>To enable a fair visual comparison of training dynamics, each curve is normalized as in <a href="http://www.plosone.org/article/info:doi/10.1371/journal.pone.0335135#pone.0335135.e243" target="_blank">Eq (29)</a> by dividing by its epoch-1 value and EMA-smoot...
שמור ב:
| מחבר ראשי: | |
|---|---|
| מחברים אחרים: | , , |
| יצא לאור: |
2025
|
| נושאים: | |
| תגים: |
הוספת תג
אין תגיות, היה/י הראשונ/ה לתייג את הרשומה!
|
| סיכום: | <p>To enable a fair visual comparison of training dynamics, each curve is normalized as in <a href="http://www.plosone.org/article/info:doi/10.1371/journal.pone.0335135#pone.0335135.e243" target="_blank">Eq (29)</a> by dividing by its epoch-1 value and EMA-smoothed (exponential moving average) with . The EMA is computed as with . This normalization emphasizes relative convergence behavior (shape and stability) rather than raw magnitudes: with RLC, the total objective continues to decrease in late epochs, whereas without RLC it plateaus, consistent with the ablation trends in <a href="http://www.plosone.org/article/info:doi/10.1371/journal.pone.0335135#pone.0335135.t005" target="_blank">Table 5</a>.</p> |
|---|