Speaker
Description
Keywords: Neuromorphic-Computing, spiking neural network, computational neuroscience, HPC, Edge Computing
Brain-inspired Spiking Neural Networks represent a promising frontier in computational models, offering potential advantages over traditional computing paradigms in terms of energy efficiency, temporal information processing, and adaptability to dynamic data. This can benefit numerous applications, such as real-time signal processing and pattern recognition in resource-constrained environments. The research landscape in neuromorphic computing, which encompasses hardware architectures designed to efficiently implement these biologically-inspired networks, is highly heterogeneous, with diverse approaches balancing biological plausibility against computational efficiency. In this diversity there is significant opportunity for exploration of novel architecture designs and applications.
This presentation introduces our work on a new multi-core neuromorphic architecture prototype that is under development within the INFN Brainstain project, where we have brought together the diverse expertise present inside CSN5, from the design and implementation of high performance computing architectures dedicated to physics tasks, to the modeling of novel neuron models that enable incremental learning through efficient integration of contextual and sensory information and the mechanism of apical amplification.
Leveraging on the proprietary APEIRON framework for flexible, low latency communication we aim to deploy our architecture prototype on a multi-FPGA system. We adopt a software-hardware codesign workflow that relies on early validation through a high level simulator of the architecture, and on the High Level Synthesis (HLS) programming paradigm for translation to a hardware implementation, for relatively fast and simple reprogramming, debugging and feature enhancements. The flexibility of this approach and the modular design of our architecture will allow us to explore support for different models of neuron dynamics, such as multi-compartment neuron models, and study the system performance with different inter-core communication schemes, such as specialized broadcast or multicast algorithms.
This presentation will discuss our architectural approach, the features and capabilities planned for the system, it will describe the current status of development, and it will outline the future direction of the project.
