CODEWITHSUNDEEP
pythontensorflowneural-networkartificial-intelligencekeras
Asdf11
Asdf11

Reputation: 447

Keras regression multiple outputs

a question concerning keras regression with multiple outputs:

Could you explain the difference beteween this net:

two inputs -> two outputs

input = Input(shape=(2,), name='bla')
hidden = Dense(hidden, activation='tanh', name='bla')(input)
output = Dense(2, activation='tanh', name='bla')(hidden)

and: two single inputs -> two single outputs:

input = Input(shape=(2,), name='speed_input')
hidden = Dense(hidden_dim, activation='tanh', name='hidden')(input)
output = Dense(1, activation='tanh', name='bla')(hidden)

input_2 = Input(shape=(1,), name='angle_input')
hidden_2 = Dense(hidden_dim, activation='tanh', name='hidden')(input_2)
output_2 = Dense(1, activation='tanh', name='bla')(hidden_2)

model = Model(inputs=[speed_input, angle_input], outputs=[speed_output, angle_output])

They behave very similar. Other when I completly seperate them, then the two nets behave like they re supposed to.

And is it normal that two single output nets behave much more intelligible than a bigger one with two outputs, I didnt think the difference could be huge like I experienced.

Thanks a lot :)

Upvotes: 4

Views: 5989

Answers (1)

Mach_Zero
Mach_Zero

Reputation: 524

This goes back to how neural networks operate. In your first model, each hidden neuron receives 2 input values (as it is a 'Dense' layer, the input propagates to every neuron). In your second model, you have twice as many neurons, but each of these only receive either speed_input or angle_input, and only works with that data instead of the entire data.

So, if speed_input and angle_input are 2 completely unrelated attributes, you'll likely see better performance from splitting the 2 models, since neurons aren't receiving what is basically noise input (they don't know that your outputs correspond to your inputs, they can only try to optimize your loss function). Essentially, you're creating 2 seperate models.

But in most cases, you want to feed the model relevant attributes that combine to draw a prediction. So splitting the model then would not make sense, since you're just stripping neccessary information away.

Upvotes: 4

Related Questions