search

djanloo / cmepda

An LSTM + Encoder network for UHECR AirShowers
1 stars 2 forks source link

Biased estimation ? #6

Closed djanloo closed 2 years ago

djanloo commented 2 years ago

The distribution of true - norm. estimated seems to have a linear trend.

The normalization operation requires the knowledge of the true value, but a linear trend should be the easiest to learn. Why doesn't it seem to be so? interpercentile_ranges

djanloo commented 2 years ago

The net does not have the true value as a reference while doing predictions. The linear trend has a deeper origin, it cannot be learned by a 1-unit dense layer. I am dumb.