The deployment process of a spiking neural network (SNN) can involve partitioning a neural network and mapping partitions onto processing units within the neuromorphic hardware. Searching for optimal deployment schemes presents an NP-hard problem. Optimization of deployment schemes encounters challenges in devising computationally effective cost functions for optimization objectives such as communication time consumption and energy efficiency. These kinds of objectives necessitate consideration of network dynamics shaped by neuron activity patterns, demanding intricate mathematical analyses or simulations for integrating them into a cost model for the deployment of an SNN. The network dynamics are hardware-independent and can be modeled separately from specific hardware configurations. Our approach employs a pairwise Ising-type maximum entropy model, which has shown its effectiveness in accurately reproducing pairwise correlations among components in a system. We utilized this model to capture network dynamics, upon which a cost function is built incorporating hardware-specific parameters. We conducted an extremely preliminary investigation using the SpiNNaker machine. We show that the existing model training can also be computationally complex. Currently, we still lack sufficient evidence to substantiate the effectiveness of our proposed methods. Further efforts is needed to explore integrating network dynamics into SNN deployment.

翻译：脉冲神经网络（SNN）的部署过程可能涉及对神经网络进行分区，并将分区映射到神经形态硬件内的处理单元上。寻找最优部署方案是一个NP难问题。部署方案的优化面临着设计计算高效的成本函数的挑战，这些函数用于优化目标，如通信时间消耗和能源效率。这类目标需要考虑由神经元活动模式塑造的网络动态，这需要复杂的数学分析或模拟才能将其集成到SNN部署的成本模型中。网络动态与硬件无关，可以独立于特定的硬件配置进行建模。我们的方法采用了一种成对伊辛型最大熵模型，该模型已证明能有效准确地再现系统中组件间的成对相关性。我们利用该模型捕捉网络动态，并在此基础上构建了一个包含硬件特定参数的成本函数。我们使用SpiNNaker机器进行了极其初步的探索。我们发现，现有模型的训练也可能计算复杂。目前，我们仍缺乏足够的证据来证实所提方法的有效性。需要进一步努力探索将网络动态集成到SNN部署中。