Hardmaru @ Tokyo University

@hardmaru

scott mccloud tom white, sampling generative networks (2016)

BEGAN ( a GAN that works like a VAE, in that you can interpolate all the latent vectors )

(x, y, P-touch, P-lift, P-end) as a row of data

RNN -> latent vector -> add gaussian noise to latent vector -> decode RNN

World models

• purpose: to understand how ml models learn abstract representations of the world

echo state networks?

otoro.net

creatures avoiding planks

• with random nn weights
• if they touch a tank they die
• if they bump into another agent, they grow little neural networks

worldmodels.github designrl.github