Use Beta NLL for regression when you also predict standard deviations, a simple change to NLL that works reliably better.

Freiburg

Tübingen

Imagine a task where you give a list of edges of a star graph, start and end node, and train a model with a teacher forcing you to predict the list of tokens in the path from the start to the end.

(edge 1, edge 2 ...) (start, goal) (start, intermediate1, intermediate 2 .. .goal)

V/LM AR glasses always had this Rick and Morty death crystals vibe to me.

LMAO 🤣

blueskyroast.com/roast/saipra...

Done :D

Arxiv sharing reminder

pdf ❌
abs ✅

Event poster made with so much creativity and love!

Klammer Zwei: A Local Freiburg Jazz band, killing it 🔥

The cRSSM qualitatively extrapolates to unseen context values in terms of generated image samples. When presented a counterfactual cF, the decoded images from imagined latent states is conditioned on the counterfactual, showing disentanglement of the infered latent states.

We find the cRSSM consistently outperforming the following three settings,
hidden: Dreamer trained on multiple contexts without the context being observable.
concat: Concatenating context with observations.
default: Dreamer trained only on default context.

Towards ZSG we introduce the novel contextual recurrent state-space model (cRSSM). It incorporates context into the world model of DreamerV3, allowing it to infer latent states which are (perhaps) disentangled from the context and also to predict latent dynamics/dream of many worlds.

The context comprises of parameters of a POMDP such as mass/length of a robot. It affects the dynamics and unlike (latent) states such as position/velocity, context does not change during an episode. We aim for ZSG to contexts in- and out-of-distribution.

I haven't used the Semanticscholar pdf reader a lot, but knew it has a highlighter. But now they have AI based auto-highlights!! Seems really rad, I will start using it in my workflow and report back if it's useful as I start to use it!