Philipp Becker

Excited to announce that today we will present four papers at the ICML workshops : 📐 @NiklasFreymuth will showcase “Iterative Sizing Field Prediction for Adaptive Mesh Generation From Expert Demonstrations” @ the AI for Science Workshop. Stay tuned! #ICML2024 #MachineLearning

@sai_prasanna @artemZholus @apsarathchandar @danijarh Thanks for tagging me. I think @artemZholus makes a good point in that the graphical model induces an inductive bias and thus it depends on the problem. In my experience, using tighter lower bounds is more important when there is a lot of uncertainty in the system.

@artemZholus @apsarathchandar @philippb06 Also @danijarh , do you see any practical issues when making representation model non-recurrent (the variational posterior on latents only depend on observations) to make training parallel with S4? I find it hard to see how we can derive a posterior only with observations

🎉 Excited to announce our work "Recall to Imagine (R2I)" got an oral spot at #ICLR2024! R2I is a generalist and computationally efficient agent that shines in RL memory tasks, showing superhuman performance in most complex memory tasks. Kudos to @M_R_Samsami & @artemZholus! 1/n

🧩 How does #BayesianDeepLearning handle the wild terrain of real-world o.o.d. data? Novel work from our student Florian Seligmann with the #WILDS 🐾 datasets: Exploring transformers, finetuning and last-layer methods while ensembling it all! #NeurIPS2023 arxiv.org/abs/2306.12306

Feeling the stress of slow simulations?⏰Let Adaptive Swarm Mesh Refinement (ASMR) soothe your computational woes! Our method produces high-quality mesh refinements that offer up to 100x speedup.🎧📊 arxiv.org/abs/2304.00818 github.com/NiklasFreymuth… Catch us at #NeurIPS2023!

I'm super excited about this PhD work of mine, "Multi Time Scale World Models", which has been accepted to Neurips 2023 as a spotlight (Top 3% of all submitted papers). Details are in the thread below. (1/6)

For more details, check out our paper: openreview.net/forum?id=UQXdQ… , Code available at github.com/pbecker93/vrkn . Many thanks also to the reviewers, action editor, and organizers of @TmlrOrg . We really like the new format!

As a more principled alternative, we propose the VRKN. It uses a smoothing approach for an inference that considers future observations and explicitly models epistemic uncertainty. Thus, it gives correct aleatoric uncertainty estimates and improves performance in many cases.

I am happy to share our recent work, “On Uncertainty in Deep State Space Models for Model-Based Reinforcement Learning”, ( openreview.net/pdf?id=UQXdQyo… ) published at TMLR. With @geri_neumann Let me summarize our key contributions in a short thread:

Excited to present our work on "Hidden Parameter RSSMs for Changing Dynamics Scenarios" at ICLR 2022 on 27 April. 🙂 We learn a deep multi-task Kalman Filter than can adapt to non-stationary environments. openreview.net/forum?id=ds8yZ… @geri_neumann @dtrbchlr @philippb06 @alr_lab_kit

Our work ‚Specializing Versatile Skill Libraries using Local Mixture of Experts‘ is accepted at #CoRL2021 🙂. openreview.net/forum?id=j3Rgu… Thanks to Prof. Neumann(@alr_kit), Zhou, Ge Li and Philipp Becker (@philippb06).

Our work 'Action Conditional Recurrent Kalman Network' is accepted at #CoRL2020. Excited to share more details at the virtual conf. :) Thanks to Prof Gerhard Neuman(@alr_kit) and our collaborators @philippb06, @dtrbchlr, @MarcHanheide, @LCAS_UoL. arxiv.org/abs/2010.10201

@MLSS_Tuebingen @MPI_IS Is Maximum Likelihood the right objective for multi-modal data? Especially versatile behavior? And how can we efficiently use alternative objectives? youtube.com/watch?v=vUQ2_k…

Posters @ #MLSS2020 To participants, please link your posters' videos in this thread! Would be great to include a brief intro as well :) @MLSS_Tuebingen @MPI_IS