2022 BRN Discussion, page-2787

Thebelow documents commence with a paper explaining Graph Neural Networks.

Thesecond is an extract and link to a paper suggesting the benefits that can beobtained by integrating a spiking neural network to make Graph Neural Networkspower efficient and practical for edge deployment.

Thethird is a press release from Socionext claiming they have achieved a GraphNeural Network that achieves “1/60th of the time needed by conventionaltechnologies. The new method enables advanced SLAM processing even on SoCs foredge devices with limited CPU performance and power consumption”.

https://bdtechtalks.com/2021/10/11/what-is-graph-neural-network/amp/

https://openreview.net/forum?id=Ul3o26VB6KZ

Keywords: Graph, spike,energy, neural network

Abstract: Graph ConvolutionalNetworks (GCNs) achieve an impressive performance due to the remarkablerepresentation ability in learning the graph information. However, GCNs, whenimplemented on a deep network, require expensive computation power, which makesthem difficult to be deployed on battery-powered devices. In contrast, SpikingNeural Networks (SNNs), which perform a bio-fidelity inference process, offeran energy-efficient neural architecture. In this work, we propose SpikingGCN,an end-to-end framework that aims to integrate the embedding of GCNs with thebiofidelity characteristics of SNNs. In particular, the original graph data areencoded into spike trains based on the incorporation of graph convolution. Wefurther model biological information processing by utilizing a fully connectedlayer combined with neuron nodes. In a wide range of scenarios, includingcitation networks, image graph classification, and recommender systems, ourexperimental results show that the proposed method could gain competitiveperformance against state-of-art (SOTA) approaches. Furthermore, we show thatSpikingGCN on a neuromorphic chip can bring a clear advantage of energyefficiency into graph data analysis, which demonstrates its great potential toconstruct environment-friendly machine learning models

https://socionextus.com/pressreleases/socionext-and-tohoku-university-significantly-accelerate-deep-learning-based-slam-processing/

No opinion orspeculation but Socionext has access to the most advanced SNN edge processor inthe World.

FF

AKIDA BALLISTA