Auto-encoder As close as possible NN NN Encoder code Decoder Randomly generate a vector as code code NN Decoder Image
Auto-encoder As close as possible NN Encoder NN Decoder code NN Decoder code Randomly generate a vector as code Image ?
Review:Auto-encoder 2D code NN Decoder 9 NN Decoder -1.50 01.5 NN Decoder
Review: Auto-encoder NN Decoder code 2D -1.5 1.5 −1.5 0 NN Decoder 1.5 0 NN Decoder
Review:Auto-encoder 5 555 3 3 3 5 5 0 3 2 8 2 -1.5 1.5 0 0 a 00 0 0 99
Review: Auto-encoder -1.5 1.5
Auto-encoder NN NN input output Encoder Decoder code VAE Minimize reconstruction error NN 3 input C1 NN Encoder 01 output exp Decoder 2 C3 ci=exp(o)×e+m From a normal distribution 123 Minimize 3 (exp(c)-(1+o)+(m)2)
NN Encoder NN Decoder code input output Auto-encoder VAE NN Encoder input NN Decoder output m1 m2 m3 𝜎1 𝜎2 𝜎3 𝑒3 𝑒1 𝑒2 From a normal distribution 𝑐3 𝑐1 𝑐2 X + Minimize reconstruction error 𝑖=1 3 𝑒𝑥𝑝 𝜎𝑖 − 1 + 𝜎𝑖 + 𝑚𝑖 2 exp 𝑐𝑖 = 𝑒𝑥𝑝 𝜎𝑖 × 𝑒𝑖 + 𝑚𝑖 Minimize
What will happen if we only Why VAE? minimize reconstruction error? Original Intuitive Reason Code Code with noise input NN m3 C1 NN Encoder 01 xp C2 output Decoder 2 C3 03 e1 3 Minimize exp(ci)-(1+0)+(m)2） The variance of noise is automatically learned
Why VAE? Intuitive Reason NN Encoder input NN Decoder output m1 m2 m3 𝜎1 𝜎2 𝜎3 𝑒3 𝑒1 𝑒2 𝑐3 𝑐1 𝑐2 X + exp Original Code Code with noise The variance of noise is automatically learned What will happen if we only minimize reconstruction error? 𝑖=1 3 𝑒𝑥𝑝 𝜎𝑖 − 1 + 𝜎𝑖 + 𝑚𝑖 2 Minimize