使用BP神经网络进行深度学习时训练时获得平坦的误差曲线

Question

I am always getting flat curves for error plots while deep learning with conventional BP neural networks. I am using Keras sequential model with Adam optimiser. The NN giving overall 80% accuracy both for training and testing. Can anyone explain why the error curves are flat (see attached figure)? Also is there any way to improve my results?

def build_model():
  model = keras.Sequential()
  model.add(layers.Dense(128, activation=tf.nn.relu, input_shape=len(normed_train_data.keys())]))
  model.add(layers.Dense(128,activation=tf.nn.relu, input_shape=(1,)))
  model.add(layers.Dense(4))
  model.compile(loss='mean_squared_error', optimizer='Adam',metrics=['mae', 'mse','accuracy'])
  return model

def plot_history(history):
   hist = pd.DataFrame(history.history)
   hist['epoch'] = history.epoch
   plt.figure()
   plt.xlabel('Epoch')
   plt.ylabel('Mean Abs Error [per]')
   plt.plot(hist['epoch'], hist['mean_absolute_error'],label='Train Error')
   plt.plot(hist['epoch'], hist['val_mean_absolute_error'],label = 'Val Error')
   plt.legend()
   plt.ylim([0,200])
   plt.show()

And in the main function,

model = build_model()
model.summary()
history = model.fit(normed_train_data, train_labels,epochs=EPOCHS,validation_split = 0.2, verbose=0,callbacks=[PrintDot()])
hist = pd.DataFrame(history.history)
hist['epoch'] = history.epoch
plot_history(history)

error plots :

Error plot with reduced learning rate

Answer 1

It is difficult to assess that without having more information about your data, can you share a sample? But I'd hazard a guess that your model overfits very quickly. Things that you can try:

• model simplification -- try removing one layer, or using less units for starters
• different optimizer, try sgd with different learning rates
• different metrics (try removing one by one)