variational-autoencoders #

Shows the VAE pipeline as an animated, blocky graph: an input x is encoded into variational parameters (μ_φ(x), σ_φ(x)) for q_φ(z|x), sampled via the reparameterization z=μ+σ⊙ε, then decoded through p_θ(x|z) to generate x~. A bottom panel animates the ELBO/KL decomposition as a stacked bar to emphasize that maximizing ELBO tightens the bound by reducing the KL gap.

canvasclick to interact

⏮◀◀▶▶STEP0.25x1xZOOM

t=0s

practical uses #

01.Generate new samples similar to training data (images, audio, embeddings)
02.Learn compressed latent representations for clustering/interpolation
03.Semi-supervised learning and anomaly detection via reconstruction + KL regularization

technical notes #

Pure Canvas2D; responsive scaling via scale=min(w,h)/240 and snapped grid g for a retro aesthetic. Animation cycles ~4.2s through encoder→reparam→decoder→ELBO focus. A small particle system (closure state) visualizes ε flowing into z then into decoder output, reinforcing the reparameterization trick and gradient path.

← basis-and-dimension predicate-logic →