Autoencoders & GANs by Hand: Encoders, Bottlenecks, Reconstruction Error, Anomaly Scores, Latent Space, Variational Sampling, Generator vs Discriminator Losses & Mode Collapse
Open up the two classic generative networks. You will squeeze tiny toy images through a bottleneck and decode them, measure reconstruction error, take one training step, flag anomalies, walk through a latent space, sample a variational code with given noise and a KL term, score real and fake samples with a discriminator and read a generator–discriminator training log for mode collapse, with given weights and every cell checked as you go.
One-time payment · lifetime access · lessons, labs, exam and certificate included · no subscription.
About this course
This course opens up autoencoders and GANs with 2 × 2 toy images and given weights. You encode images into a two-value bottleneck, count parameters, decode with sigmoid outputs and measure squared reconstruction error. You take one gradient step on the decoder weights and show the loss drop. You turn reconstruction error into an anomaly score with a percentile threshold and grade it with precision and recall. You interpolate in latent space, clean a noisy input, sample variational codes with a given ε table, compute the KL term and weigh it with β. Finally you score real and generated samples with a discriminator, calculate both GAN losses and check a training run for mode collapse.
Lessons
- The Compression Idea: 2 × 2 Toy Images, an Encoder and a Two-Value Bottleneck Pro
- Decoding and Reconstruction Error: Sigmoid Outputs and Squared Errors Pro
- One Training Step: Gradients for the Decoder Weights, a Learning Rate and the Loss Before and After Pro
- Anomaly Scores: Reconstruction Errors, a Percentile Threshold, Precision and Recall Pro
- Latent Space: Codes on a Chart, Interpolating Between Images and Denoising Pro
- Variational Autoencoders: Mean and Log-Variance, Sampling with Given ε, the KL Term and β Pro
- GANs – the Two Players: Discriminator Scores, Discriminator Loss and Generator Loss Pro
- Capstone – Reading a Training Run: the Balance Point, Spread, Occupied Bins and Mode Collapse Pro