Fig. 2.

Schematic overview of the proposed convolutional neural network architecture. The network receives two inputs: an image and the treatment indicator ($t$). Loss functions are depicted in double octagons. The last layer activations are used to separate factors of variation in the image. $a_1$ is trained to approximate the measurement of the collider $x^{\prime }$. The rest of the last layer activations are constrained to be linearly independent from $x^{\prime }$ through $L_{\mathrm{reg}}$. The total loss is $L=L_y+L_x+L_{\mathrm{reg}}$. CNN convolutional neural network.