Investigating normalising convolutional/non-convolutional features

chengsoonong / crowdastro

Cross-identification of radio objects and host galaxies by applying machine learning on crowdsourced training labels.

MIT License

13 stars 1 forks source link

Investigating normalising convolutional/non-convolutional features #228

Closed MatthewJA closed 7 years ago

MatthewJA commented 7 years ago

Doesn't seem to change very much — changes seemed smaller than the deviation from subsets. Curiously, if I weighted the non-convolutional features higher during training, and then didn't weight during testing, the classifier performed a little bit better, but this is probably noise.