Unsupervised Deep Learning¶

Traditional Autoencoder¶

Feature learning, dimensionality reduction, anomaly detection

flowchart LR
x1["x"] -->
|Encoder| z["z"] -->
|Decoder| x2["x̂"]

• Usually \(\vert z \vert < \vert x \vert\), to find useful small subset of features
• Sometimes encoder and decoder share weights
• Use encoder to initialize a supervised model

Error function will be \(u_i = x_i - \hat x_i\)

Variational Autoencoder¶

• Bayesian
• Useful to generate new data

GAN¶

Generative Adversarial Networks

flowchart LR
n[/Noise/] ---> g[Generator] --> d
rd[Real Data] -->
d[Discriminator] -->
rf{Real/Fake} -.->
|Backpropagation| d & g

Multiscale¶

Vector Math¶

Comments