Below is keras records of training model, epoch=20, no matter how much I modify model always around 12th epoch test set, the phenomenon of the loss significantly higher, what reason is this excuse me? I use the loss function is MSE, may for the impulse noise had magnified effect, but the question is why not every epoch of the MSE is very high, but the training to the 12th epoch when suddenly increased?

Epoch 1/20
60/60 [==============================] - 167 - s 3 s/step - loss: 0.8877 - val_loss: 0.2337
Epoch 2/20
60/60 [==============================] - 166 - s 3 s/step - loss: 0.7081 - val_loss: 0.1796
Epoch 3/20
60/60 [==============================] - 166 - s 3 s/step - loss: 0.6973 - val_loss: 0.1650
Epoch 4/20
60/60 [==============================] - 168 - s 3 s/step - loss: 0.6907 - val_loss: 0.0897
Epoch 5/20
60/60 [==============================] - 167 - s 3 s/step - loss: 0.6847 - val_loss: 0.1025
Epoch 6/20
60/60 [==============================] - 164 - s 3 s/step - loss: 0.6809 - val_loss: 0.1731
Epoch 7/20
60/60 [==============================] - 167 - s 3 s/step - loss: 0.6770 - val_loss: 0.2273
Epoch 8/20
60/60 [==============================] - 167 - s 3 s/step - loss: 0.6744 - val_loss: 0.1221
Epoch 9/20
60/60 [==============================] - 167 - s 3 s/step - loss: 0.6724 - val_loss: 0.5085
Epoch 10/20
60/60 [==============================] - 166 - s 3 s/step - loss: 0.6713 - val_loss: 0.4322
Epoch 11/20
60/60 [==============================] - 165 - s 3 s/step - loss: 0.6695 - val_loss: 0.2419
Epoch 12/20
60/60 [==============================] - 167 - s 3 s/step - loss: 0.7354 - val_loss: 11.4644
Epoch 13/20
60/60 [==============================] - 165 - s 3 s/step - loss: 0.6667 - val_loss: 4.5552
Epoch 14/20
60/60 [==============================] - 164 - s 3 s/step - loss: 0.6653 - val_loss: 0.0955
Epoch 15/20
60/60 [==============================] - 165 - s 3 s/step - loss: 0.6629 - val_loss: 0.0924
Epoch 16/20
60/60 [==============================] - 165 - s 3 s/step - loss: 0.6596 - val_loss: 0.1004
Epoch 17/20
60/60 [==============================] - 167 - s 3 s/step - loss: 0.6558 - val_loss: 0.3345
Epoch 18/20
60/60 [==============================] - 167 - s 3 s/step - loss: 0.7868 - val_loss: 0.3541
Epoch 19/20
60/60 [==============================] - 168 - s 3 s/step - loss: 0.6708 - val_loss: 0.2062
Epoch 20/20
60/60 [==============================] - 166 - s 3 s/step - loss: 0.6507 - val_loss: 0.1870