3
$\begingroup$

I'm playing around with TCN's lately and I don't understand one thing. How is the receptive field different from the input size?

I think that the receptive field is the time window that TCN considers during the prediction, so I guess the input size shall be equal to it.

According to the WaveNet paper, I cannot see a reason why it should be otherwise. I'm using TensorFlow with this custom library.

Please help me understand.

$\endgroup$
2
$\begingroup$

As you can see in Fig. 2 of the WaveNet paper the receptive field is 5, but the input size is larger (16). The receptive field defines what a single output neuron can see (see arrows in Fig. 2).

enter image description here

The receptive field could also be greater than the input, e.g. if you want to use or you only have the last 12 time steps and use the following structure (WaveNet paper, Fig 3), which can cover different powers of two depending on the number of layers.

enter image description here

If you want to calculate not only the last output neuron, it can make sense that the input size is larger than the receptive field, as the outputs before use also older inputs (see dashed lines).

$\endgroup$
3
  • 1
    $\begingroup$ Thank You a lot for the answer. So if I look only at one neuron output it would make sense to match input size with receptive field? Also if I conclude correctly, I could put another output neuron on the top of TCN that would take several TCN's outputs as an input and in such way make my predictions more reliable? $\endgroup$ Nov 16 '21 at 7:16
  • $\begingroup$ 1) Yes, if you only use the last output neuron, it makes sense to match the input size with the receptive field. 2) No, that shouldn't actually do much. The individual output neurons correspond to the individual time steps. Combining the outputs would not produce much more information. If you want to cover a larger time horizon, you should rather adjust the other parameters to increase the receptive field instead. $\endgroup$
    – dexteritas
    Nov 16 '21 at 8:23
  • $\begingroup$ OK, thank You very much. $\endgroup$ Nov 16 '21 at 8:55

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.