Generating new data with VAEs

hamish - 29/05/2020

• bit like the "identity function" of NNs
• aim is to learn the input via a small "bottleneck"
• break this task up into an "encoder" and "decoder"

Autowhat now?

(Autoencoder)

https://www.i2tutorials.com/explain-about-convolutional-autoencoder/

Latent Space

Can we force the latent space to be more consistent?

Add noise before bottleneck

• forces VAE to compensate for noise
• does this by learning to place similar things close to one another
• different things far apart
• more consistent space

How does this work?

https://towardsdatascience.com/intuitively-understanding-variational-autoencoders-1bfe67eb5daf

"Shift" in a direction

http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html

https://houxianxu.github.io/assets/project/dfcvae

1. find direction of feature we like
2. use encoder to find position of image in latent space
3. move in feature direction
4. decode from that point

Downside: Implementing this is surprisingly hard, even with the code

So why do we care?

• We (CIA) want to be able to provide actions so people can lower their energy use
• Would like to be able to "simulate" scenarios like if user got an EV

• not really learnt the peaks you get with EVs
• causes issues with standby
• lose strong signal from other devices

David Ha https://worldmodels.github.io/