SNNs Robustness

Learning to Predict Memory Robustness from Spiking Neural Networks.

dataset

Load dataset

import pickle

dataset = pickle.load(open('RCN_dataset.pickle', 'rb'))

dataset['features'] # axis labels (metrics)
dataset['X'] 		# 400 datapoints (15D vectors)
dataset['Y'] 		# 400 target values (maximal connections drops)

Features

0. nonrecurrent_count: number of non-reciprocal edges (a->b but not b->a).
1. recurrent_count: number of reciprocal edges (a->b and b->a).
2. cliques_count: number of maximal cliques within the graph.
3. k_edge_connect: k value at which the graph becomes disconnected.
4. clique_size_avg: average maximal clique size (number of nodes).
5. clique_size_std: standard deviation of clique_size_avg.
6. cs_max: biggest maximal clique.
7. in_degree_centrality_avg: average in-degree centrality of nodes.
8. in_degree_centrality_std: standard deviation of in_degree_centrality_avg.
9. out_degree_centrality_avg: average out-degree centrality of nodes.
10. out_degree_centrality_std: standard deviation of out_degree_centrality_avg.
11. between_centrality_avg: average betweenness-centrality of nodes.
12. between_centrality_std: standard deviation of between_centrality_avg.
13. closeness_centrality_avg: average closeness-centrality of nodes.
14. closeness_centrality_std: standard deviation of closeness_centrality_avg.

Observations

1. Some metrics (e.g. cliques count) are integers while others (e.g. in/out-degree centrality) are continuous values - normalization might be needed.
2. Some data points (the ones with target values very low or very high) might be underrepresented - filtering might be needed.