What Machine Learning Looks Like

This is what training a convolutional neural network to remove the stars from astronomical photographs looks like. Neural networks are amazing, but there is nothing magical or "intelligent" about them. In essence, they are sophisticated mathematical functions that can "learn" to do various tasks through training. It requires an intelligence at the keyboard to define them, coax them through training, and most importantly to define the task they are to perform.
Technical details: the network trained in this video is a convolutional U-Net with DenseNet and dilated convolution elements, comprising about 21M parameters to train. Training took approximately 18,000 iterations through a heavily-augmented data set to ensure generalization. The loss function is a combination of mean absolute, mean squared, and log-mean-exponential errors.

[@cosgrovescosmos]

This is just fascinating! What a great way to illustrate the learning process!

[@RaviAdhav]

Me too

[@luke6575]

Mentour pilot bought me here