The authors propose simmering, a physics-inspired alternative to optimization-based neural network training that generates weights through systematic sampling rather than optimization, to mitigate overfitting and achieve better generalization compared to conventional methods.
- Irina Babayan
- Hazhir Aliahmadi
- Greg van Anders
