Fig. 2: Classification of molecular graphs.
From: Echo state graph neural networks with analogue random resistive memory arrays
a, An illustration of some samples from the MUTAG molecular dataset, where nodes of different colours represent different atoms while edges are chemical bonds. Depending on the mutagenicity, these molecules are categorized into positive and negative classes. b, An example MUTAG node embedding process. The input features of all nodes, defined as X, are first projected onto the state space using the input matrix \({{{{W}}}}_{{{I}}}\), and the hidden state of each node is updated according to the protocol shown in Fig. 1f and Methods, which leads to node embeddings that encapsulate graph information. c, The graph embedding vectors of the two categories of the MUTAG dataset. Each column vector is a graph embedding. The embeddings of the left (right) colour map are from the positive (negative) class. d, The graph embeddings are mapped to a 2D space using PCA. Pink (blue) dots represent molecules with positive (negative) mutagenicity, which can be linearly separated. e, The accuracy of each fold in a ten-fold cross-validation and the software baseline. The average accuracy is 92.11%, comparable to state-of-the-art algorithms. f, The confusion matrices of the experimental classification results. The upper matrix is a ten-fold averaged confusion matrix, which is then normalized horizontally to produce the lower matrix. g, A breakdown of the estimated MAC OPs (red bars) and associated energy (light-blue bars for a state-of-the-art GPU; dark-blue bars for a projected random resistive memory-based hybrid analogue–digital system). In a forward (backward) pass, the fully optimized model on a state-of-the-art GPU and the ESGNN on a projected random resistive memory-based hybrid analogue–digital system consume approximately 74.32 µJ (approximately 160.51 MOPs) and approximately 34.44 μJ (approximately 1.05 MOPs), respectively, revealing a >2.16 fold improvement of the inference energy efficiency (a 99.35% reduction of the backward pass complexity). BP, backpropagation.
