[Python] Why is my neural network overfitting even after using dropout and L2 regularization?

I’m working on a binary classification problem using a neural network in TensorFlow/Keras.
The dataset has around 9,500 training samples and 2,500 testing samples.

My model architecture:

model = Sequential([
Dense(128, activation='relu', input_shape=(X_train.shape[1],)),
Dropout(0.4),
Dense(64, activation='relu', kernel_regularizer=tf.keras.regularizers.l2(0.001)),
Dropout(0.3),
Dense(1, activation='sigmoid')
])


I’m using:

• 
  Optimizer: Adam
  
  
• 
  Loss: Binary Crossentropy
  
  
• 
  Epochs: 30
  
  
• 
  Batch size: 32
  

Even after adding dropout and L2 regularization, the training accuracy keeps going up (≈97%) while the validation accuracy stays around (≈81%). Validation loss is also increasing, which clearly looks like overfitting — but I’m already using regularization techniques.

Continue reading...