D-EMA
A parameter regularization technique that applies stronger constraints to interference-sensitive parameters, such as late-layer MLPs, through grouped EMA while preserving the flexibility of other parameters, used to mitigate parameter interference in multi-task learning.
Timeline (last 90 days)
DynThinker is a unified framework aimed at jointly strengthening static and dynamic spatial reasoning, whose core consists of two complementary components, D-EMA and SDF.
Decoupled exponential moving average (D-EMA) adopts a configurable grouped averaging strategy that constrains interference-sensitive pretrained parameters while preserving flexibility for other parameters.
All Facts (2)
DynThinker is a unified framework aimed at jointly strengthening static and dynamic spatial reasoning, whose core consists of two complementary components, D-EMA and SDF.
50%UnverifiedDecoupled exponential moving average (D-EMA) adopts a configurable grouped averaging strategy that constrains interference-sensitive pretrained parameters while preserving flexibility for other parameters.
50%