Massively Parallel Implementation Of Sparse Message Retrieval Algorithms In Clustered Clique Networks

Auto-associative memories are a family of algorithms designed for pattern completion. Many of them are based on neural networks, as is the case for Clustered Clique Networks which display competitive pattern retrieval abilities. A sparse variant of these networks was formerly introduced which enables further improved performances. Specific pattern retrieval algorithms have been proposed for this model, such as the Global-Winners-Take-All and the Global-Losers-Kicked-Out. We hereby present accelerated implementations of these strategies on graphical processing units (GPU). These schemes reach interesting factors of acceleration while preserving the retrieval performance.