A neural network trained on collisions of protons and electrons transferred its skills to a completely different experiment — catching neutrinos. Like a chef who learned French cuisine and then, without preparation, baked an excellent Uzbek plov.
Researchers took the OmniLearned model, trained on data from particle accelerators where energies are enormous and detectors are complex. Then they tested it on events from MINERvA — a detector that catches neutrinos from an accelerator with energies of just a few gigaelectronvolts. The neural network had to guess particle energies and recognize whether pions were born in the reaction.
The result surprised: the model, which had never seen neutrino data, performed better than its twins trained from scratch specifically for MINERvA. With the same amount of computation and the same number of training steps, the transferred model gave more accurate answers. This means that during training on diverse data, the neural network learned general patterns of particle physics that work at other energies, detectors, and interaction types.
Such knowledge transfer promises a revolution in data analysis: instead of training a new model for each specific experiment, one can use a single universal tool. This will save time and computational resources, and also speed up discoveries.
And here is the chef who mastered French cuisine already cooking plov, and doing it better than local chefs. Perhaps soon physicists will have a universal assistant that understands the language of particles regardless of where and how they were born.
Neutrinos are so elusive that every second trillions of these particles pass through your body, but in a lifetime only one or two will interact with the matter of your body.