1 paper · 1 filter
Shivam Agarwal, Zimin Zhang, Lifan Yuan +2
Entropy minimization (EM) trains the model to concentrate even more probability mass on its most confident outputs. We show that this simple objective alone, without any labeled da…