Thread by @danijarh

Excited to present Clockwork VAEs for video prediction!

Clockwork VAEs (CW-VAEs) leverage hierarchies of latent sequences, where higher levels tick slower. They learn long-term deps across 1000 frames, semantically separate content, and outperform strong video models.

👇 Thread

On 4 diverse datasets, Clockwork VAEs yield more accurate long-term predictions than strong baselines. VTA also uses temporal abstraction but its consecutive frames are inconsistent. RSSM and SVG make plausible predictions but ignore long deps. SVG sometimes falls off manifold.

We perform a comprehensive empirical evaluation, comparing Clockwork VAE to a hierarchy ticking at the input rate (NoTmpAbs) and to the previous video models RSSM, SVG, and VTA. On average over 4 datasets and 4 metrics, Clockwork VAE achieves the best rank of the five methods.

More specifically, Clockwork VAEs are hierarchies of latent sequences, where each latent conditions multiple latents at the level below. Each state consists of the recurrent state of a GRU and the stochastic latent variable that incorporates bottom-up information about the input.

We find a direct correspondence between the temporal abstraction factor — how much slower each level ticks — and how far Clockwork VAEs accurately predict ahead. Higher factors not only improve long-term predictions but also reduce computation and speed up training.

To see what's stored at each level, we only allow the bottom and middle levels to look at the context input. Top follows the prior. We find global info (room colors) is lost and thus was kept at the top, while details at lower levels (position of camera and nearby walls) remain.

Clockwork VAEs automatically adapt to the frame rate of the input sequence. The KL loss at each level gives a bound on how much information is stored at the level. The faster objects in the video move, the more the top level is used and the less middle and lower levels are used.

Congrats to Vaibhav Saxena (@saxenavaibhav11), who did a fantastic job on this! He's on the PhD market this year and has my highest recomm. Check out our paper and the website with videos and the Minecraft dataset.

Paper: arxiv.org/pdf/2102.09532…
Website: danijar.com/cwvae