search

ErikOsinga / master_research_project

Repository for the Master Research Project: "Data compression for weak lensing studies with the upcoming Euclid mission with Information Maximizing Neural Networks"
0 stars 0 forks source link

test whether the summaries are correlated or not! #8

Open mjvakili opened 5 years ago

mjvakili commented 5 years ago

Can you look at the output of the initial IMNN training and see whether {xi} parameters are correlated or not?

ErikOsinga commented 5 years ago

Yes, they seem to be correlated

The correlation matrix is the following: [[ 1. -0.2900731] [-0.2900731 1. ]]

x1vsx2

mjvakili commented 5 years ago

But didn't these points correspond to different cosmologies?

ErikOsinga commented 5 years ago

Ah yes. When I put 5000 noisy datavectors generated with the fiducial parameters into the network the output is as follows

output_fid

The output summaries are thus highly correlated:

[[1.         0.95541443]
 [0.95541443 1.        ]]
mjvakili commented 5 years ago

Nice! does that mean that the additional regularization function added to the objective function promotes perfect correlation instead of zero correlation? Can you check the code to see how the regularization look like?

Does this also imply that in order to constrain the parameters you can safely only use x1?

ErikOsinga commented 5 years ago

The regularization function seems to be implemented correctly, square_norm = tf.reduce_sum( tf.square( tf.subtract( cov, tf.eye(self.n_summaries))), name="square_norm_covariance") coupling = tf.placeholder( dtype=self._FLOATX, shape=(), name="coupling") loss = tf.subtract( tf.multiply(coupling, square_norm), logdetfisher, name="loss")

I think since \Omega_m and \sigma_8 are degenerate, it is possible to constrain the combination of the two parameters with only x1.

mjvakili commented 5 years ago

Yes. This seems reasonable. Perhaps switching the parameters to to S8 and Omegam will change it. See issue #10

ErikOsinga commented 5 years ago

Summaries are still correlated for S8 and Omega_m:

summaries_3