1
$\begingroup$

In a machine learning application, a model learns to apply a filter h(n) via convolution to a 1-dimensional input signal s(n) (e.g. sampled at Fs = 16 kHz). The filter size (filter length?) is limited to a fixed number, e.g. k = 100, so the model cannot learn a filter that exceeds this duration but it is free to insert zeros and make the filter "shorter".

Does this mean that the filter size limits the duration of the "slowest" signal component that falls into the passing bandwidth and hence Fs/k = 160 Hz is the minimal bandwidth that h(n) can represent (assuming it's a band pass filter)?

$\endgroup$
  • 1
    $\begingroup$ yes, the length of a filter limits the "frequency-selectiveness". If you remember modern physics, that's mathematically very related to the Heisenberg uncertainty principle: something that is temporally short can't be very sharp in the frequency domain, and vice versa. $\endgroup$ – Marcus Müller Jun 2 at 14:15
  • $\begingroup$ Also, filter design is one of the oldest application fields of evolutionary / ML algorithms. You'll be delighted to find that even some of the very "classical", non-ML algorithms for bandpass filter design make extensive use of a loss / reward function while optimizing a filter. If you're doing modern machine learning for filer design, there's a lot of modern literature, but I'd really start by reading up on the Remez exchange (or just Remez) algorithm, to give you a starting direction. Understanding that relatively simple algorithm and its convergence will give you a solid base! $\endgroup$ – Marcus Müller Jun 2 at 14:19
  • $\begingroup$ (especially when later writing a report on it, that would be the most basic benchmark for your ML approach that I could think of. Compare yourself to that!) $\endgroup$ – Marcus Müller Jun 2 at 14:20
  • $\begingroup$ @MarcusMüller You might want to turn your comments into an answer. $\endgroup$ – Peter K. Jun 3 at 12:40
  • 1
    $\begingroup$ @PeterK. you've got a point there. $\endgroup$ – Marcus Müller Jun 3 at 16:22
2
$\begingroup$

Yes, the length of a filter limits the "frequency-selectiveness". If you remember modern physics, that's mathematically very related to the Heisenberg uncertainty principle: something that is temporally short can't be very sharp in the frequency domain, and vice versa.

You can find one example of such a constraint in this answer, though the formula only applies to specific filter constraint.

Generally, filter design is one of the oldest application fields of evolutionary / ML algorithms. You'll be delighted to find that even some of the very "classical", non-ML algorithms for bandpass filter design make extensive use of a loss / reward function while optimizing a filter.

If you're doing modern machine learning for filer design, there's a lot of modern literature, but I'd really start by reading up on the Remez exchange (or just Remez) algorithm, to give you a starting direction. Understanding that relatively simple algorithm and its convergence will give you a solid base!

Especially when later writing a report on it, that would be the most basic benchmark for your ML approach that I could think of. Compare yourself to that!

$\endgroup$

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.