60 mice, 1,980 samples, 234 lung cancer patients. Behind each number lie thousands of genes. Their activity is measured to understand the disease. But the data aren’t uniform: for one patient, you assess a protein level (a continuous number), make a diagnosis (yes/no), and determine the tumor subtype (a category). Mixing them in a single model is like adding degrees Celsius to kilometers. Experts in Machine Learning manually tune error weights for each task. Without a common language, there's no way to agree on importance — the outcome depends on the case.

The authors proposed abandoning absolute scales and switching to a common language of order. Here’s how it works. The model, like a translator, doesn’t fixate on exact numbers. It grasps the hierarchy — who is higher, who is lower. It learns to rank patients: whose marker is larger, which tumor is more aggressive. This approach is invariant to monotonic transformations — operations that change numbers but not order, like taking a logarithm. In practice, the strict “greater/less” rule is replaced by a smooth sigmoid. This makes optimization proceed smoothly.

The neural network has a simple design. A shared first layer sifts through hundreds of gene and keeps only the universal players. A group penalty (a feature selection method that zeroes out all parameters for uninformative features) removes noise. Then, for each task, there are separate layers that output predictions in arbitrary order scores. Experiments in numerical simulation and on real bioinformatics data showed that the method outperforms dynamic weighting, GradNorm, and single-task networks.

The selected genes turned out to be no accident. Scd1 governs lipid metabolism. E2F1 is a transcription factor in breast cancer. RPLP2 is a marker for lung adenocarcinoma. The common language of order not only boosts accuracy. It translates disparate measurements into a unified biological narrative that links immunity and oncogenesis. Each data type ceases to be an isolated dialect. Those 60 mice, 1,980 samples, and 234 patients are no longer just statistics — their genomes, RNA profiles, and diagnoses now speak in one language the neural network understands.