0
$\begingroup$

I am currently creating a neural network to learn a function of the following form Data that I want to learn x corresponds to x axis and y to y axis(one dependent and one independent variable)

enter image description here

I am using both keras and tensorflow and with both scripts I get the following result

enter image description here I am currently creating a neural network to learn a function of the following form Data that I want to learn x corresponds to x axis and y to y axis(one dependent and one independent variable)

I am using both keras and tensorflow and with both scripts I get the following result Predictions orange line Data blue line. Somehow my neural network doesn't capture the non-linearity of the data and only tries to fit a linear function. Do you maybe have a suggestion what I am doing wrong? Also is the architecture appropriate for the following task or there exists some problems.

Additionally as an information I also include a snippet of the architecture that I am using in keras

    def individual_model(keys, labels, config):
model = Sequential()
model.add(Dense(32, input_dim=1))
model.add(LeakyReLU())
for i in range(2):
    # if str(i) not in config:
    #     break
    model.add(Dense(32))
    model.add(LeakyReLU())
model.add(Dense(1))
model.compile(optimizer='adam', loss='mse', metrics=[
              max_absolute_error, 'mse', 'mae'])
model.fit(keys, labels, epochs=100, batch_size=32, verbose=1) 
return model
Improve this question
$\endgroup$
2
  • $\begingroup$ Could you please add the output of model.summary() to your question? Thanks. $\endgroup$ – georg-un Jul 23 '19 at 12:06
  • $\begingroup$ You can find a good example of function approximation here. I think you can try using the sigmoid function instead of LeakyReLU. $\endgroup$ – Shubham Panchal Jul 23 '19 at 12:31
1
$\begingroup$

You should try using a sigmoid activation function, as leakyRelu is very close to a linear function, and remove one hidden layer. Just one hidden layer should be engough as you have only one input, with a very small length, maybe 3 or 5 neurons.

Improve this answer
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.