1 of 21

MPNP

{ccc53*,bjd39*,arj39,pl219}

2 of 21

Neural Processes (NPs)

3 of 21

Neural Processes (NPs)

  • NNs + GPs — best of both worlds!
  • Learn from datasets (X)
  • Represent a collection of stochastic processes fi and model uncertainty via z
  • Generative model:

4 of 21

Often we have more than just X, we have A

  • Dataset with relations (A) between points (X)
  • Molecules, social/traffic/citation networks, meshes, ...
  • Can leverage message passing (MP):

5 of 21

MPNP (Message-Passing Neural Process)

6 of 21

MPNP

  • encoder

​

  • aggregation

,

  • decoder

7 of 21

MPNP

8 of 21

Data sampling

  • context
    • information available to the encoder

​

  • target
    • information available to the decoder: similar to above

9 of 21

Fixed-class experiments

TUD : PROTEINS, ENZYMES, DHFR, COX2, BZR

ShapeNet single category

Cellular automata

10 of 21

Biochemical datasets

11 of 21

ShapeNet Single Category

12 of 21

ShapeNet Single Category

13 of 21

Cellular Automata

  • regular & irregular graph structure,
  • deterministic & probabilistic evolution
  • count & density based rule types
  • relational & geometric edges
  • planar, toroidal, spherical & non-Euclidean topologies

14 of 21

Mixed-class… more challenging - introduce concat

  • Arbitrary labelling
  • CNP Omniglot basis

15 of 21

Mixed-class tasks

Cora-Branched

ShapeNet mixed-category

16 of 21

Cora-Branched

17 of 21

Cora-Branched few-shot tasks

18 of 21

ShapeNet mixed-category

19 of 21

Mixed-category ShapeNet

20 of 21

Future work

  • modelling the generative process for structure
  • NPs known to underfit & Attentive NPs fix that → design AMPNPs!

21 of 21

Thank you for listening!