GSoC 23 Progress

Created on July 26|Last edited on September 25
Comment
Developed an efficient Ego QGNN Architecture in JAX and TensorCircuit.
﻿
﻿
Node 0:
1 hop neighbours: 1, 2, 5, 3
2 hop neighbours: 4, 6
3 hop neighbours: 7
Have to perform padding each ego-graph.
Now the shape of the dataset is (N, MaxNodes, NumHops, MaxNodes in a ego graph).
Training on MUTAG. Overfitting.
﻿
﻿
Next steps:
Explore distributed training with JAX for training on quark gluon.
﻿
Ego graphs with Quark Gluon
DatasetThe number of qubits quickly scales up with the number of k-hop neighbours.
Tested with only 1000 samples of the dataset
+----------+----------+
|   k      | # qubits |
+----------+----------|
|  1       | 6        |
|  2       | 23       |
|  3       | 41       |
|  4       | 47       |
|  5       | 54       |
+----------+----------+
Too slow with higher values of k.
Figure from pennylane blog.﻿
Hyperparameter-free and Explainable Whole Graph Embedding ﻿Paper﻿
Based on Degree, H-index and coreness and Shannon entropy. 
Can be used be data-reuploading circuit and therefore this leads to a single qubit QNN. 
	
﻿
Add a comment