Patrick van der Smagt
@smagt.bsky.social
scientist in machine learning and robotics / serious-music lover / language pedant / ethical AI champion
Can you unmake a Swiss roll?
VAEs compact but destroy data shape. Solution: add a distance-preserving loss based on a continuous kNN graph. Using VAEs with learned priors, our models have topologically consistent latent spaces.
Paper: Nutan Chen & Botond Cseke, ICONIP 2024 arxiv.org/abs/2206.05909
VAEs compact but destroy data shape. Solution: add a distance-preserving loss based on a continuous kNN graph. Using VAEs with learned priors, our models have topologically consistent latent spaces.
Paper: Nutan Chen & Botond Cseke, ICONIP 2024 arxiv.org/abs/2206.05909
December 13, 2024 at 1:42 PM
Can you unmake a Swiss roll?
VAEs compact but destroy data shape. Solution: add a distance-preserving loss based on a continuous kNN graph. Using VAEs with learned priors, our models have topologically consistent latent spaces.
Paper: Nutan Chen & Botond Cseke, ICONIP 2024 arxiv.org/abs/2206.05909
VAEs compact but destroy data shape. Solution: add a distance-preserving loss based on a continuous kNN graph. Using VAEs with learned priors, our models have topologically consistent latent spaces.
Paper: Nutan Chen & Botond Cseke, ICONIP 2024 arxiv.org/abs/2206.05909
#NeurIPS: Offline RL causes distribution shift + value overestimation. Our Constrained Latent Action Policies (C-LAP)
🔹 generative modelling of observations and actions mimics online learning
🔹 constrained latent actions let policy stay in data distribution
arxiv.org/abs/2411.04562
Dec 11 4:30pm PST
🔹 generative modelling of observations and actions mimics online learning
🔹 constrained latent actions let policy stay in data distribution
arxiv.org/abs/2411.04562
Dec 11 4:30pm PST
December 9, 2024 at 11:08 AM
#NeurIPS: Offline RL causes distribution shift + value overestimation. Our Constrained Latent Action Policies (C-LAP)
🔹 generative modelling of observations and actions mimics online learning
🔹 constrained latent actions let policy stay in data distribution
arxiv.org/abs/2411.04562
Dec 11 4:30pm PST
🔹 generative modelling of observations and actions mimics online learning
🔹 constrained latent actions let policy stay in data distribution
arxiv.org/abs/2411.04562
Dec 11 4:30pm PST