Imagine we have an 7-D function f(P_i), i=1,...,7. Fortunately, we can create our own data sets for training, i.e we can choose as many samples from the sample space of each P_i (the sample space is known), what I need to know is that how we should choose our data sets for training? (for example, a uniform random selection from the sample space of each parameter is a good choice?)
example:
as an example, I want to approximate a bi-variate Gaussian function which is the function of x1,x2 , mux1,mux2(mean of the distribution), sigmax1,sigmax2 , ro (variance of the distribution)
f(x1,x2,mux1,mux2,sigmax1,sigmax2,ro) and the sample space of each parameter is known to me, for example I know a
1. Deep embedding based data selection can be one option.
2. The one you said is also sounds ok. But too much random selection may hamper the performance.
Dear Hasan
From your answer I had this impression that There is no optimal sampling method, or at least we can not say that Deep embedding based data selection is definitely better than uniform random selection. is it correct?
So I think the point is that we should choose denser data-sets in higher dimensional spaces, for example using Halton or Sobol sequence to generate data-set help the neural network to be trained much better in comparison with choosing random data-set.
So I think the point is that we should choose denser data-sets in higher dimensional spaces.
- Badges
- Science topic
- Similar topics
- Organizational Psychology
- Training