Handwritten Text Recognition for manuscripts and early printed texts
NS - CUK Seminar: V.T.Hoang, Review on "Long Range Graph Benchmark.", NeurIPS 2022
1. Van Thuy Hoang
Dept. of Artificial Intelligence,
The Catholic University of Korea
hoangvanthuy90@gmail.com
Vijay Prakash Dwivedi et. al., NeurIPS 2022
2. 2
Long Range Information Bottleneck
Existing Graph Benchmarks
Characterizing Long Range Graph Datasets
Proposed Datasets and Tasks
Experiments and Questions
3. 3
Long Range Information Bottleneck
A MP-GNN layer aggregates information from its 1-hop neighbors to update a
node’s feature representation.
While MP-GNNs have several limitations, the so called Information Bottleneck,
that particularly impacts long range interactions.
4. 4
Long Range Information Bottleneck
If long-range information from an L-hop neighbor is needed for a task, L
number of layers are ideally required to be stacked.
With increasing L, the L-hop neighborhood grows exponentially and so does
the amount of information that needs to be encoded into one vector, see the
adjacent pictorial representation.
This brings a serious bottleneck in MP-GNNs when dealing with tasks that
requires long range information propagation.
5. 5
Existing Graph Benchmarks
Many of the existing graph learning benchmarks consist of prediction tasks
that primarily rely on local structural information rather than distant information
propagation to compute a target label or metric.
This can be observed in datasets such as ZINC, ogbg-molhiv and ogbg-
molpcba where models that rely significantly on encoding local (or, near-local)
structural information continue to be among leaderboard toppers.
6. 6
Existing Graph Benchmarks
This can be observed in datasets such as ZINC, ogbg-molhiv and ogbg-
molpcba where models that rely significantly on encoding local (or, near-local)
structural information continue to be among leaderboard toppers.
E.g.,
Zin dataset: 23 ndoes
OGBG-Molhiv: 25 nodes
….
Dianeter of graphs: ~30 in Enzymes, proteins datasets
7. 7
Characterizing Long Range Graph Datasets
Graph size:
The number of nodes in a graph is an important characteristic to determine
a long range graph dataset.
A term called problem radius which refers to the required range of
interaction between nodes for a particular problem.
The problem radius must be sufficiently large for a graph dataset if it
serves as a long range benchmark.
8. 8
Characterizing Long Range Graph Datasets
Although for real world graph datasets, the problem radius may not be exactly
quantified, this hypothetical metric would effectively be smaller if a graph has
smaller number of nodes.
Therefore, the (average) graph size of a dataset is a key property for
determine whether it could be a potential long range graph dataset
9. 9
Nature of task
The nature of task can be understood to be directly related to the problem
radius.
In broad sense, the task can be either short-range, i.e., requiring information
exchange among nodes in local or near-local neighborhood, or long-range,
where interactions are required far away from the near-local neighborhood.
For example,
in ZINC molecular regression dataset, the task is associated with counting
local structures and experimental revelations using a substructure-
counting based model by Bouritsas et al., 2020 has shown ZINC’s task
would optimally require counts of 7-length substructures.
ZINC’s regression task may thus be interpreted as a short-range task.
10. 10
Contribution of global graph structure to task
A dataset where the learning task benefits from global structural information
can be a potential long range graph dataset.
Sample representation of 3D atomic contact between distant nodes.
11. 11
Proposed LRGB Datasets and Tasks
Consider the characteristics described above to propose a collection of 5
graph learning datasets that can be used to prototype GNNs or Transformers
with long range modeling capabilities.
The table below for an overview of the datasets’ statistics:
12. 12
Proposed LRGB Datasets and Tasks
PascalVOC-SP and COCO-SP:
These are superpixel graphs based on Pascal VOC 2011 and MS COCO
image datasets respectively.
The learning task in both the datasets is node classification where each
node corresponds to a region of the image belonging to a particular class,
with respect to the original semantic segmentation labels in the respective
image datasets.
PCQM-Contact:
This dataset is based on a subset of PCQM4M dataset from OGB-LSC
where each graph corresponds to a molecular graph with explicit
hydrogens and the task, link prediction, is to predict pairs of distant nodes
that will be contacting with each other in the 3D space with a pre-defined
threshold.
13. 13
Proposed LRGB Datasets and Tasks
Sample visualization of a peptide (Left) and its molecular graph (Right).
14. 14
Experiments
Experiments using two major architecture class from the graph learning
literature:
local MP-GNNs
fully connected Graph Transformers, to establish benchmarking trends
and understand more about the datasets while also charting out some
probable challenges that requires further research.
15. 15
Proposed LRGB Datasets and Tasks
Q1: Is a local feature aggregation, modeled using MP-GNNs with fewer layers,
enough for the proposed tasks in LRGB?
Simple local MP-GNN instances perform poorly due to increased effect of
over-squashing.
Q2: Do we observe a visible separation in performance of models with
enhanced capability to capture long range interactions when compared
against local MP-GNNs on the proposed benchmark?
The baseline Transformers appeared slower to fit on COCO-SP on which
the recent GraphGPS architecture, that can model long range
dependencies, significantly outperforms MP-GNNs.
Q3: What are the challenges and future discoveries that can be facilitated by
the new benchmark?
There are different challenges revealed from our benchmarking
experiments that can be pursued for further investigation while using the
proposed datasets.