namdw/iclr2019
Summary of ICLR 2019 by ML2 members
README
Summary of ICLR 2019 by ML2 members
Representation Learning on Graphs and Manifolds
- 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
- Learning Mixed-Curvature Representations in Product Spaces
- 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
- 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
- Gaussian representation
- not closed under intersection
- Cone representation
- Box representation
- Ordered Neurons: Integrating Tree Structures into Recurrent Neural Networks
- 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.
