Neural-network playground

The diagram draws the network. A star enters as 48 numbers, its BP coefficients in the upper band and its RP coefficients in the lower one, and passes through the hidden layers to a single output, the estimated [M/H].

Each line is a weight, coloured by its sign and drawn thicker the larger it is; to stay legible, the diagram draws at most 400 lines, chosen by the size of their weights. Unless your device asks for reduced motion, dots trace the followed star through the layers, and each node's colour shows how strongly it responds to that star.

The error curve gives the mean squared difference between the network's estimates and the GSP-Spec values after each epoch, one pass through the 3,200 training stars. The validation stars never adjust the weights, so their curve shows how well the network generalises. When the validation curve flattens or turns upwards while the training curve keeps falling, the network has begun to fit noise in its training stars; an L2 penalty, which pulls the weights towards zero, usually delays that.

The scatter plot sets each validation star's estimate against its GSP-Spec value, which comes from a separate instrument on Gaia, the Radial Velocity Spectrometer. A perfect estimate would sit on the diagonal, and colour gives the star's temperature. Select a point to follow that star through the network.

Train runs until you pause it, for at most 2,000 epochs, and Step trains for one epoch. Reset starts again from the same initial weights, so two runs that differ in one setting can be compared. Changing the layers, the units or the activation builds a new network, while the learning rate, the batch size and the L2 penalty take effect at once.

The bars are the followed star's 48 input numbers. The curves are the spectra of three red giants of similar temperature and colour. Absorption lines of metals crowd the blue end of a spectrum, and the giant with [M/H] +0.25 has less blue light, relative to its mean, than the other two.