The Heidelberg spiking datasets for the systematic evaluation of spiking neural networks (1910.07407v3)

Abstract: Spiking neural networks are the basis of versatile and power-efficient information processing in the brain. Although we currently lack a detailed understanding of how these networks compute, recently developed optimization techniques allow us to instantiate increasingly complex functional spiking neural networks in-silico. These methods hold the promise to build more efficient non-von-Neumann computing hardware and will offer new vistas in the quest of unraveling brain circuit function. To accelerate the development of such methods, objective ways to compare their performance are indispensable. Presently, however, there are no widely accepted means for comparing the computational performance of spiking neural networks. To address this issue, we introduce two spike-based classification datasets, broadly applicable to benchmark both software and neuromorphic hardware implementations of spiking neural networks. To accomplish this, we developed a general audio-to-spiking conversion procedure inspired by neurophysiology. Further, we applied this conversion to an existing and a novel speech dataset. The latter is the free, high-fidelity, and word-level aligned Heidelberg digit dataset that we created specifically for this study. By training a range of conventional and spiking classifiers, we show that leveraging spike timing information within these datasets is essential for good classification accuracy. These results serve as the first reference for future performance comparisons of spiking neural networks.

Systematic Evaluation of Spiking Neural Networks Using Heidelberg Spiking Datasets

The paper "The Heidelberg spiking datasets for the systematic evaluation of spiking neural networks" presents a significant step towards providing standardized benchmarks for the evaluation of Spiking Neural Networks (SNNs). SNNs are computational models inspired by the workings of the biological brain, notable for their energy efficiency, parallel information processing capabilities, and noise tolerance. Despite advancements in instantiating SNNs for practical applications, until now, the field has lacked widely accepted benchmarks for comparing the performance across various SNN architectures and learning algorithms.

Introduction to Benchmarking Challenges

SNN architectures vary significantly, and the methods for training these networks can be diverse and complex. Conventional benchmarks like the MNIST dataset serve the traditional Artificial Neural Network (ANN) community but require adaptations for SNNs, such as conversion to spikes, which undermines comparability due to the variability in transformation processes employed by different research groups.

Benchmark datasets offer quantitative methods for comparing different approaches, fostering competitive advancements in the field. The authors address critical gaps by introducing two spike-based classification datasets developed through the audio-to-spiking conversion methods inspired by neurophysiology: the Heidelberg Digits (HD) and the Spiking Heidelberg Digits (SHD).

The Heidelberg Spiking Datasets

1. Heidelberg Digits Dataset (HD): This novel dataset comprises high-fidelity recordings of spoken digits, specifically constructed for spiking neural network research. The dataset uses audio recordings processed through a biological model of the inner ear, yielding spikes that are directly applicable to SNN models, bypassing the need for manual spike conversion. This promotes consistency and objectivity in performance evaluations across varied SNN implementations.
2. Spiking Heidelberg Digits Dataset (SHD): The SHD variant applies audio-to-spiking conversion processes to existing and novel datasets of spoken commands, facilitating benchmarking for keyword spotting beyond simple digit recognition. These datasets enable the community to evaluate the efficacy of spike timing information in classification tasks, an area where SNNs provide significant advantages over conventional ANNs.

Methodological Framework

The conversion methodology involves spatial frequency dispersion of the audio signals using a hydrodynamic basilar membrane model. Subsequent models of hair cells transform frequency data into spike trains, attributing biologically plausible characteristics to the spikes generated. These spikes form the inputs to SNNs specifically designed and trained within this paper, which demonstrate the impact of spike timing in achieving high classification accuracy.

Results and Observations

The paper establishes baseline performances using both linear and non-linear classifiers, including Support Vector Machines (SVMs) and Convolutional Neural Networks (CNNs), illustrating the necessity of spike timing for nuanced classification tasks. The authors assert that while a linear SVM does not generalize well to SHD data, LSTM models achieve significant accuracy, showcasing the advantage of recurrent structures in processing temporally coded data.

Implications and Future Directions

The introduction of these datasets represents a pivotal move towards standardizing SNN evaluation methodologies. By fostering a shared benchmarking platform, this work is poised to accelerate scientific understanding and practical development of SNNs for diverse applications. The datasets enable researchers to objectively measure improvements in efficiency and accuracy, fueling further advancement in neuromorphic computing and non-von-Neumann architectures. Future research could explore the integration of additional sensory modalities, broadening the applicability of SNN-driven models.

In summary, the Heidelberg spiking datasets are a vital resource for systematic evaluation across SNN architectures, encouraging standardized comparative analyses that can drive innovation in brain-inspired computational models. These contributions lay the groundwork for exploring more sophisticated and functionally relevant SNN applications, with implications spanning both theoretical neuroscience and practical AI developments.