Mobile Intelligence

17.2 GRAPH NEURON

The GN is a novel approach that uses graph-based representations of patterns to achieve one-shot learning. A highly-scalable associative memory device is thus created, which is capable of handling multiple streams of input that are processed and matched with the historical data stored within the network. The method uses parallel in-network processing to circumvent the pattern-database scalability limitation associated with graph-based techniques [8].

17.2.1 Associative Memory Concept

Associative memory (AM) is derived from the neural network model and has been applied in many different application areas [5]. A widely used unsupervised learning technique is the Hopfield network. This network has been widely used for implementing associative (or content-addressable) memory in pattern analysis and optimization. A study of Hopfield memory model shows that the model is not scalable and is limited by the number of processing/storage nodes in the network [9]. The backpropagation network provides fast recalls, but the training cost becomes excessive for adding newer patterns. Ideally, an associative memory device should include simple one-shot training in the case of discrete time processing and fast retrieval. The GN is an approach that aims to overcome the scalability issues and reduce the training overheads in associative memory devices [10].

17.2.2 Simple Graph Neuron Approach

The GN is a finely distributed in-network pattern-recognition algorithm that preserves the ...