À suivreGuide suivant
Réseaux de neurones convolutifs
Fondamentaux
GUIDE des fondamentaux
Scaling laws are empirical relationships between language-model loss and quantities such as parameters, training data, and compute.
Écoutez ce guideNarré par l'IA · ~2 min
They can help researchers plan experiments and extrapolate within studied regimes, but they do not guarantee an exact result for a new architecture, dataset, or training setup.
Kaplan and colleagues’ 2020 study found empirical power-law relationships between language-model cross-entropy loss and model size, dataset size, and training compute over the regimes they tested. On log-log plots, a power law appears approximately linear. These fitted relationships can be used to estimate outcomes for larger runs, but extrapolations are conditional on comparable training and data setups. The original paper found large models to be more sample-efficient in its experiments and suggested compute-optimal training with relatively large models and fewer tokens than later Chinchilla work recommended. Later studies, including Llama 3, explored training beyond that earlier compute-optimal allocation.
Fitting a loss curve from smaller controlled runs to inform a larger training-budget decision
Comparing parameter and data allocations under a fixed pretraining-compute budget
Checking whether a forecast remains accurate when the data mixture or architecture changes
Separating measured loss trends from downstream capability claims in a model report
Scaling-law research studies new data mixtures, modalities, architectures, and inference-time computation. As training and data regimes change, researchers must remeasure the relevant curves. Scaling laws support resource planning, while downstream accuracy, safety, and reliability require separate evaluation.
Newsletter gratuite
Trois histoires d'IA vérifiées chaque matin de la semaine, écrites dans un anglais simple. Gratuit pour toujours, sans publicité.
Un e-mail chaque jour de la semaine. Désabonnez-vous en un clic. Nous ne vendons ni ne partageons jamais votre adresse.
Testez-vous
Commentaires instantanés sur chaque réponse et un certificat partageable avec une pièce d'identité vérifiable une fois que vous avez réussi un cours.
Soutenez l’enseignement gratuit de l’IA. AI Understanding est une organisation à but non lucratif 501(c)(3) : pas de publicité, pas de paywall, jamais. Faire un don
Scaling laws are empirical relationships between language-model loss and quantities such as parameters, training data, and compute. They can help researchers plan experiments and extrapolate within studied regimes, but they do not guarantee an exact result for a new architecture, dataset, or training setup.
Kaplan et al. report empirical power-law relationships over their tested ranges; the fit should not be read as a guarantee outside comparable settings.
Kaplan et coll. étudié la perte en tant que loi de puissance dans le nombre de paramètres (N), les jetons d'entraînement (D) et le calcul total (C).
Fitted curves can inform large-run forecasts from smaller experiments, but changes in data, architecture, or training conditions can limit extrapolation.
La perte comporte une partie réductible qui diminue avec l'échelle, plus un plancher irréductible fixé par le caractère aléatoire inhérent aux données.
With L(X)=E+A·X^-α, a fixed multiplicative increase gives a similar fractional change in the reducible term, but smaller absolute changes as that term shrinks.
Continuez à apprendre
Plus de guides sélectionnés pour ce sujet
À suivreGuide suivant
Réseaux de neurones convolutifs
Fondamentaux