Validation loss doesn't decrease（驗證 loss 不會下降）

[AdrienGuo]

This image is the line plot of loss. Solid line is the loss of train, dashed line is the loss of validation. The loss of validation doesn't decrease. However, the weird thing is, regardless the loss of validation doesn't decrease. The model of 50th epoch is clearly better than the 1st epoch.

Does anyone know this weird phenomenon?

================================================ 下圖是我用 GOT-10K 當作資料集的 loss 曲線。 實線是訓練的 loss，虛線是驗證的 loss，驗證 loss 很明顯沒有在下降。 不過奇怪的是，儘管驗證的 loss 沒有下降，但第50個epoch的model 很明顯比第 1 個還要好很多。

有人可以解釋這種現象嗎？