Authors - Gabriel M. da Silva, Nicolas O. da Rocha, Heloise V. C. Brito, Joao V. N. M. da Silva, Sergio A. S. da Silva, Anderson R. de Souza, Carlos A. O. de Freitas, Vandermi Joao da Silva Abstract - Spiking Neural Networks (SNNs) have been investigated as a biologically inspired alternative for efficient information processing, particularly in energy-sensitive applications. This work presents a comparative evaluation of the energy efficiency of different SNN techniques, including Liquid State Machines (LSM), Recurrent Spiking Neural Networks (RSNN), Spiking Convolutional Neural Networks (SCNN), and learning based on Spike-Timing Dependent Plasticity (STDP). The experiments were conducted on conventional hardware plat-forms, namely an Android smartphone and a notebook, using simulated implementations of SNNs without dedicated neuromorphic acceleration. The analysis considered different network scales by varying the number of neurons and was based on neural activity metrics, particularly the total number of generated spikes, employed as a proxy for the indirect estimation of energy consumption during audio signal processing. The results demonstrate a consistent relationship between neural activity and estimated energy consumption, as well as an energy saturation behavior as network complexity increases. Differences among the an-alyzed techniques are more pronounced in small-scale configurations, whereas larger networks exhibit convergent patterns of neural activity and energy consumption. Although conducted in a digital simulation environment, this study highlights the limitations of conventional platforms for the efficient execution of SNNs and reinforces the potential of dedicated neuromorphic hardware for embedded and low-power applications.
Open Zoom