Architecture of our ensemble network, which produces 4 outcomes as the end product – to be exact, the probability of a candidate being a lensed quasar, a galaxy, an unlensed quasar, and a star – is shown here. The input layer takes the HSC grizy-band images and passes them to the functional layers containing 12 CNN and 2 ViT models. Each model produces 4 output logits, which are then transformed into a probability distribution that sums to unity using the softmax layer. Subsequently, we average the softmax outputs of these 14 networks to produce only 4 final probabilities.