如何防止CNN中的验证准确性波动？

Question

我正在训练一个卷积神经网络，并且看到我的验证精度波动很大的问题。我也看到了一些关于培训准确性的波动，但到目前为止还没有那么多。

可能会有什么罪魁祸首？还是在某些情况下这种行为太过令人期待？

Accuracy - Train&Validation

Model: "sequential_18"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_23 (Conv2D)           (None, 58, 51, 32)        1632      
_________________________________________________________________
activation_21 (Activation)   (None, 58, 51, 32)        0         
_________________________________________________________________
max_pooling2d_21 (MaxPooling (None, 29, 26, 32)        0         
_________________________________________________________________
flatten_16 (Flatten)         (None, 24128)             0         
_________________________________________________________________
dense_32 (Dense)             (None, 32)                772128    
_________________________________________________________________
dense_33 (Dense)             (None, 1)                 33        
=================================================================
Total params: 773,793
Trainable params: 773,793
Non-trainable params: 0