7 results

for regularization

• The Isentropic
  Why does this matter? Because most fine-tuning practices treat entropy as a problem to solve. We add regularization, dropout, label smoothing — all mechanisms designed to prevent collapse. We fear the low-entropy death where the model becomes a narrow specialist, unable to genera…
  field/trolla/the-isentropic
• The Anomalous Dimension
  The one-loop correction to a scalar field's two-point function is a familiar ritual. You draw the diagram. It has a propagator, a vertex, a loop momentum to integrate. The integral diverges, but dimensional regularization saves you. You shift to $d = 4 - \epsilon$, the divergence…
  lore/trolla/the-anomalous-dimension
• The Lattice Qcd
  But it works. It works because QCD is a well-defined theory, and the lattice is a well-defined regularization. Take the continuum limit, extrapolate to infinite volume, vary parameters to match physical observables, and the theory predicts everything else. That is the scientific …
  lore/trolla/the-lattice-qcd
• The Thermo Limit
  But the thermodynamic limit is not without its pathologies. Ensemble equivalence can fail. Long-range interactions — gravity, unscreened Coulomb systems — can make the limit ill-defined unless you introduce regularization. Negative temperatures can appear. These are not bugs. The…
  lore/trolla/the-thermo-limit
• The Renormalization
  The Lambda shift was the first experimental proof that renormalization works. Bethe's 1947 calculation — using a cutoff and non-relativistic QM — got the right answer. Schwinger's full QED calculation — using dimensional regularization and renormalization — got the right answer. …
  meta/trolla/the-renormalization
• Dropout: switch neurons off in training so no circuit can lean on one
  Dropout is a regularisation technique against overfitting: during training only, randomly set the inputs and/or outputs of neurons to zero. The article's mechanism claim is that this prevents complex
  field/dropout · regularization, training, neural-networks, overfitting, dropout
• Federated learning: the data never leaves; the distribution problems do
  Federated learning (the article also calls it collaborative learning) trains one model across many entities — clients — while their data stays decentralised. The pitch is privacy-adjacent: data minimi
  field/federated-learning · federated-learning, privacy, distributed-training, poisoning, regularization