Potential Fourier Analysis Metrics?

[X89codered89X]

So, my friend looked at this post and told me it's beyond confusing. So let me clarify.

Suppose I have a neural network connected to various sensors. How best would I process the input data from the sensor such that a neural network could learn from it best. I'm assuming my network has many, many inputs, so perhaps I could input all types of processed input from my sensors. Forget about processing power limitations for now. I'm simply asking, what types of frequency input would be useful for a neural network to learn from?

I thought of one metric. I was considering weighting recent inputs higher in a Fourier integral as below.Continuous:

${\bf{F}}_c(j\omega,\alpha,t) = \int^t_{t_0} f(\tau)e^{(\tau -t)\alpha}e^{-j\omega\tau}d\tau, \; \alpha >0$

and Discrete (which I would use on a comp):

${\bf{F}}_d(j\omega,\alpha,n) = T\sum^n_{k=k_0} f_ke^{[k-n]\alpha T}e^{-j\omega Tk}, \; \alpha >0$

The idea is that they are weighted Fourier Transforms such that the exponential terms in the integrand and summand are such that since $\tau -t \leq 0, k-n \leq 0$ ,then $e^{(\tau -t)\alpha} \leq 1, e^{(k -n)T\alpha} \leq 1$ so that most recent terms are weighted exponentially more.

Then, on the network inputs, I could input values of my weighted Fourier transforms from several values of time and several frequency values, so perhaps a whole matrix of inputs from this frequency metric.

I could even do the same thing for the continuous time inputs

Continous:
${\bf{f}}_c(\alpha,t) = f(t)e^{(\tau -t)\alpha}$

Discrete:
${\bf{f}}_d(\alpha,n) = f_ne^{[k-n]\alpha T}$

and then, I can get a vector of inputs if I input the last so many sample values from this metric

Would this be a good "feature selection"?

 

[anorlunda]

Wow. That's a big topic. Start here https://en.wikipedia.org/wiki/Deep_learning

 

[osilmag]

There are labs out there that are already attempting to do something very similar to what you are writing about. Analyzing a neural signal with the Fourier Transform is nothing new. How I envisioned back in the day is that each different network responds to a different frequency. So after the Fourier Analysis of a neural network, you would build a filter (input and output) specific to that part of the network. Plenty of example code out there for that.