Neuromorphic computing shows promise for advancing computing efficiency and capabilities of AI applications using brain-inspired principles. However, the neuromorphic research field currently lacks standardized benchmarks, making it difficult to accurately measure technological advancements, compare performance with conventional methods, and identify promising future research directions. This article presents NeuroBench, a benchmark framework for neuromorphic algorithms and systems, which is collaboratively designed from an open community of researchers across industry and academia. NeuroBench introduces a common set of tools and systematic methodology for inclusive benchmark measurement, delivering an objective reference framework for quantifying neuromorphic approaches in both hardware-independent and hardware-dependent settings. For latest project updates, visit the project website (neurobench.ai).

We propose a spiking neural network (SNN) approach for radar-based hand gesture recognition (HGR), using frequency modulated continuous wave (FMCW) millimeter-wave radar. After pre-processing the range-Doppler or micro-Doppler radar signal, we use a signal-to-spike conversion scheme that encodes radar Doppler maps into spike trains. The spike trains are fed into a spiking recurrent neural network, a liquid state machine (LSM). The readout spike signal from the SNN is then used as input for different classifiers for comparison, including logistic regression, random forest, and support vector machine (SVM). Using liquid state machines of less than 1000 neurons, we achieve better than state-of-the-art results on two publicly available reference datasets, reaching over 98% accuracy on 10-fold cross-validation for both data sets.

Spiking neural network algorithms require fine-tuned neuromorphic hardware to increase their effectiveness. Such hardware, mainly digital, is typically built on mature silicon nodes. Future artificial intelligence applications will demand the execution of tasks with increasing complexity and over timescales spanning several decades. The multi-timescale requirements for certain tasks cannot be attained effectively enough through the existing silicon-based solutions. Indium-Gallium-Zinc-Oxide thin-film transistors can alleviate the timescale-related shortcomings of silicon platforms thanks to their bellow atto-ampere leakage currents. These small currents enable wide timescale ranges, far beyond what has been feasible through various emerging technologies. Here we have estimated and exploited these low leakage currents to create a multi-timescale neuron that integrates information spanning a range of 7 orders of magnitude and assessed its advantages in larger networks. The multi-timescale ability of this neuron can be utilized together with silicon to create hybrid spiking neural networks capable of effectively executing more complex tasks than their single-technology counterparts.

We have analysed unidirectional delay traces of a diverse set of IPv6 microflows routed over W-LAN and W-WAN environments. Using a number of time-domain and frequency-domain estimators we have examined the existence and intensity of long-range dependence in packet delay when viewed as time-series data. The correlation structures of packet delay on bulk TCP data path and UDP flows follow asymptotic decay while Hurst exponent estimates suggest from moderate to strong intensity (H → 1).