Skip to content

Graphs

Event-based graph neural network layer for spike-graph message passing.

  • StochasticGraphLayer — GNN convolution where message passing happens via bitstreams. Takes adjacency matrix + per-node feature vectors, propagates through graph structure with SC arithmetic. Supports variable-topology graphs.
Python
import numpy as np
from sc_neurocore.graphs import StochasticGraphLayer

adj = (np.random.rand(20, 20) > 0.7).astype(float)
np.fill_diagonal(adj, 0)
layer = StochasticGraphLayer(adj_matrix=adj, n_features=16)
features = np.random.rand(20, 16)
output = layer.forward(features)

sc_neurocore.graphs.gnn

StochasticGraphLayer

Event-Based Graph Convolution Layer. Message Passing happens via Bitstreams.

Source code in src/sc_neurocore/graphs/gnn.py
Python
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
class StochasticGraphLayer:
    """
    Event-Based Graph Convolution Layer.
    Message Passing happens via Bitstreams.
    """

    def __init__(self, adj_matrix: np.ndarray[Any, Any], n_features: int):
        self.adj = adj_matrix  # (N, N)
        self.n_nodes = adj_matrix.shape[0]
        self.n_features = n_features
        # Weights: (Features, Features) - simple linear transform
        self.weights = np.random.uniform(0, 1, (n_features, n_features))

    def forward(self, node_features: np.ndarray[Any, Any]) -> np.ndarray[Any, Any]:
        """
        node_features: (N, Features)
        """
        output = np.zeros_like(node_features)

        # 1. Message Passing (Aggregation)
        # For each node, sum neighbor features
        # In SC, this is MUX aggregation

        # Standard GCN: A * X * W

        # Aggregation:
        agg_features = np.dot(self.adj, node_features)
        # Normalize by degree? (Simplified)
        degrees = np.sum(self.adj, axis=1, keepdims=True)
        degrees[degrees == 0] = 1
        agg_features /= degrees

        # 2. Transformation (Linear)
        # Out = Agg * W
        output = np.dot(agg_features, self.weights)

        # 3. Non-linearity (Tanh/Sigmoid)
        return np.tanh(output)

forward(node_features)

node_features: (N, Features)

Source code in src/sc_neurocore/graphs/gnn.py
Python
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
def forward(self, node_features: np.ndarray[Any, Any]) -> np.ndarray[Any, Any]:
    """
    node_features: (N, Features)
    """
    output = np.zeros_like(node_features)

    # 1. Message Passing (Aggregation)
    # For each node, sum neighbor features
    # In SC, this is MUX aggregation

    # Standard GCN: A * X * W

    # Aggregation:
    agg_features = np.dot(self.adj, node_features)
    # Normalize by degree? (Simplified)
    degrees = np.sum(self.adj, axis=1, keepdims=True)
    degrees[degrees == 0] = 1
    agg_features /= degrees

    # 2. Transformation (Linear)
    # Out = Agg * W
    output = np.dot(agg_features, self.weights)

    # 3. Non-linearity (Tanh/Sigmoid)
    return np.tanh(output)