Reparameterization Trick
A core technique in VAE training that decomposes stochastic sampling into the deterministic computation z = μ + σ × ε with externally drawn noise ε. This formulation allows gradients to propagate backward through the sampling operation via the chain rule, enabling end-to-end training.
Timeline (last 90 days)
Understanding the mathematical forms of forward diffusion and reverse generation (such as noise schedules and the reparameterization trick) is a prerequisite for understanding the loss functions and sampling loops in the source code.
The reparameterization trick, proposed by Kingma and Welling in 2013, enables gradients to backpropagate through the sampling step by transforming the stochastic sampling operation into a deterministic computation plus noise.
All Facts (3)
Understanding the mathematical forms of forward diffusion and reverse generation (such as noise schedules and the reparameterization trick) is a prerequisite for understanding the loss functions and sampling loops in the source code.
50%UnverifiedThe reparameterization trick, proposed by Kingma and Welling in 2013, enables gradients to backpropagate through the sampling step by transforming the stochastic sampling operation into a deterministic computation plus noise.
50%Unverified重参数化技巧将采样表示为 z = μ + σ × ε（ε从标准正态分布采样），使随机性外置于ε，从而让μ和σ的梯度可正常反向传播实现端到端训练
50%