With just 2 labeled points, we can anchor each class to a single prototype—so the embedding updates reflect both similarity to the few labels and constraint satisfaction.

This helps in making the NeSy models #robust wrt the number of unsupervised datapoints!

🔍 Watch how the decision boundaries evolve!

This GIF shows PCA projections of the Semantic Loss model’s logits during training on MNIST-EvenOdd.

Colors = model predictions, points = true digit concepts

Above: standard semantic loss
Below: prototypical semantic loss

Want to use your favourite #NeSy model but afraid of the reasoning shortcuts?🫣

Fear not💪🏻In our #NeurIPS2025 paper we show that you just need to equip your favourite NeSy model with prototypical networks and the reasoning shortcuts will be a problem of the past!

🚨New at #ICLR: we introduce the first ever 𝐥𝐚𝐲𝐞𝐫 that makes 𝐚𝐧𝐲 neural network 𝐜𝐨𝐦𝐩𝐥𝐢𝐚𝐧𝐭 𝐛𝐲 𝐝𝐞𝐬𝐢𝐠𝐧 with constraints expressed as 𝐝𝐢𝐬𝐣𝐮𝐧𝐜𝐭𝐢𝐨𝐧𝐬 𝐨𝐟 𝐥𝐢𝐧𝐞𝐚𝐫 𝐢𝐧𝐞𝐪𝐮𝐚𝐥𝐢𝐭𝐢𝐞𝐬—even if they define 𝐧𝐨𝐧-𝐜𝐨𝐧𝐯𝐞𝐱 𝐬𝐩𝐚𝐜𝐞𝐬!