Putting a Network Under

Timing, Dose and Division of Transient Activation Suppression in Early Training

A paper from the School of Underlying Conditions applies a reversible gain to every hidden activation of a small three-layer network for a bounded window of early training, varying the onset, level, duration and division of the exposure across 1,020 runs on a synthetic task. An eight-epoch exposure at half gain costs up to 7.4 points of final accuracy, level and duration do not trade at equal area, and the lasting deficit arises at emergence, when compensated weights meet restored activations and about a third of the hidden units fall silent; restoring the gain over two epochs removes nearly all of it.

@misc{slop_5gw0bw,
  author       = {Brynjar Obiora and Ragnhild Uzoma and Vigdis Emenike},
  title        = {Putting a Network Under: Timing, Dose and Division of Transient Activation Suppression in Early Training},
  year         = {2026},
  publisher    = {Slop University},
  doi          = {10.5555/slop.5gw0bw},
  url          = {https://slop.university/outputs/slop-paper-putting-a-network-5gw0bw/},
  version      = {1.0},
  note         = {Research paper},
}