• https://rlgm.github.io/
• More robust benchmark needed.
• Geometric DL for moleular surfaces
• Many people are using GraphSAGE.
• DGL vs. PyTorch Geometry
• Le Song
  • probabilistic logic over knowledge graph
  • counter-example of GNN not being able to represent MLN exists, but can augment the GNN (for example variational GNN by Le Song) to do the job.
  • UW-CSE dataset
• Leskovec
  • Why graph hard?
    • large
    • non-unique representations of the same graph
    • long-range depedenecies
  • Graph generation baselines
    • Kronecker, MMSB, B-A

Model-agnostic meta learning

• Evolved policy gradients
• Simulation using learned model for enhanced efficiency
• Few-shot RL
• Model-based RL
  • MB-MPO (model-based meta-policy optimizations
    • Ensemble of learned simulated environments

Learn task as content in latent space Z and use the term as condition to policy Model z-space variationally using information bottleneck Compared to ProMP, MAML, RL2

• Why used information bottleneck for modeling z-space?

• Chris burges (unsupervised object …)

• Learn a good representation
• Transporter: learning object keypoints
  • Use parallel architecture with cnn and PointNet
• Object instance discovery

• (?) Wayne, Hunag, …, Lilicrap 2018
• PETES - Chua, Calandra, …, Levine 2018
• (PlaNet)Latent dynamics for pixels - Hafner, Lillicrap…, Davidson 2018

• Real world data vs generated simulation data
  • Which is better?
  • We use real world data because it has enough variation and more accurate. Generation itself is a difficult problem
• Model free vs model based model
  • What is the difference?
  • There is a big gray zone in between

• Learning Mixed-Curvature Representations in Product Spaces
  • https://openreview.net/forum?id=HJxeWnCcF7
  • 
• PyTorch Expo
  • 
  • PyTorch BigGraph
    • 
  • BoTorch
    • Highly specialized for Bayesian optimization.
    • Not for Bayesian inference therefore no altervative for TFP or Pyro.
    • 1st class support for GPyTorch models.

• Deterministic Variational Inference for Robust Bayesian Neural Networks
  • deterministic VI + empirical ELBO
    • What is empirical ELBO?
• FFJORD
  • uses neural ODE
  • continuous instead of discrete
  • can evolve a unimodal Gaussian into a multimodal complex distribution.
    • But not by a convolution with a static kernel, instead uses time evolution.
  • https://github.com/rtqichen/ffjord

• seq2tree
  • In fact, AST is also a tree from a sequence of programming language.
  • Linguists don't like sequences, but like trees.
• Embedding representation
  • point
    • no hierarchy
  • Gaussian representation
    • not closed under intersection
  • Cone representation
    • not disjoint
  • Box representation
    • Smoothing the Geometry of Probabilistic Box Embeddings
    • https://openreview.net/forum?id=H1xSNiRcF7
    • how to relate this to the ball representation on the boundary of a hyperbolic space?
• Ordered Neurons: Integrating Tree Structures into Recurrent Neural Networks
  • https://openreview.net/forum?id=B1l6qiR5F7
  • representing a branch of a tree as a stack of nodes
  • cumax
• Poincare GloVe
  • 
  • https://openreview.net/forum?id=Ske5r3AqK7
  • relation between Gauissian embedding and half-plane model hyperbolic embedding
    • mapping fixed by an isometry
    • x <-> \mu, y <-> \Sigma
    • semantic meaning by parallel transport, analogous to Euclidean vector arithmetic in a Euclidean word embedding.