A continuaciónSiguiente guía
Descenso de gradiente
Fundamentos
GUÍA DE FUNDAMENTOS
Double descent is an observed pattern in which test error can rise near the point where a model first interpolates the training data, then fall again as capacity increases.
Escuche esta guíaNarrado por IA · ~2 min
It shows that added model complexity does not always worsen test performance, though the curve depends on the model, data, and training setup.
Classical statistics teaches a U-shaped curve: as model complexity rises, test error falls, bottoms out, then rises as the model overfits. Double descent, popularized by Belkin, Hsu, Ma, and Mandal in 2019 and studied at scale by OpenAI, shows the curve has a second descent. Test error peaks near the interpolation threshold — the point where the model has just enough parameters to fit every training point exactly (zero training error). Push past that into the overparameterized regime and test error can fall again, sometimes below the classical sweet spot. OpenAI reported the phenomenon in convolutional neural networks, ResNets, and transformers; the shape and location varied by experiment. Related double-descent behavior can also appear as training duration or dataset size changes, so it is not a guarantee that every larger model will improve.
Studying how test error changes as model capacity passes the point of fitting the training set
Measuring whether training duration produces a second rise and fall in test error for a particular setup
Comparing models and datasets where increasing training-set size has different effects near interpolation
Using validation curves to understand a model’s behavior rather than assuming that a larger model always generalizes worse
Research continues to connect double descent with model class, effective complexity, implicit regularization, and optimization. The phenomenon can guide experiments on model size and training duration, but it does not prescribe that practitioners should always train larger models or longer. Validation under the target data and deployment conditions remains necessary.
Boletín gratuito
Tres noticias verificadas sobre IA cada mañana de lunes a viernes, explicadas con claridad. Gratis para siempre y sin anuncios.
Un correo cada día laborable. Date de baja con un clic. Nunca vendemos ni compartimos tu dirección.
Ponte a prueba
Corrección inmediata en cada respuesta y un certificado para compartir, con un ID verificable, cuando apruebas un curso.
Apoya la educación gratuita sobre IA. AI Understanding es una organización sin fines de lucro 501(c)(3): sin anuncios ni muros de pago, nunca. Hacer una donación
Double descent is an observed pattern in which test error can rise near the point where a model first interpolates the training data, then fall again as capacity increases. It shows that added model complexity does not always worsen test performance, though the curve depends on the model, data, and training setup.
In reported double-descent experiments, test error peaks near the interpolation threshold, where training error first reaches zero; the location and shape depend on the setup.
En la prueba del régimen sobreparametrizado, el error desciende por segunda vez, que es la característica definitoria que da nombre al doble descenso.
A specific analysis of single-hidden-layer linear networks proves a minimum-norm result under its gradient-flow and initialization assumptions; this is not a general result for all neural networks.
Puede ocurrir un doble descenso a lo largo del eje tiempo de entrenamiento: el error de prueba puede empeorar y luego mejorar a medida que el entrenamiento continúa durante más épocas.
Contradice la curva en forma de U de los libros de texto que dice que una mayor complejidad más allá de un punto siempre aumenta el error de prueba debido al sobreajuste.
sigue aprendiendo
Más guías seleccionadas para este tema.
A continuaciónSiguiente guía
Descenso de gradiente
Fundamentos