NeuroMorse: A Temporally Structured Dataset For Neuromorphic Computing

Neuromorphic engineering aims to advance computing by mimicking the brain’s efficient processing, where data is encoded as asynchronous temporal events. This eliminates the need for a synchronisation clock and minimises power consumption when no data is present. However, many benchmarks for neuromorphic algorithms primarily focus on spatial features, neglecting the temporal dynamics that are inherent to most sequence-based tasks. This gap may lead to evaluations that fail to fully capture the unique strengths and characteristics of neuromorphic systems. In this paper, we present NeuroMorse, a temporally structured dataset designed for benchmarking neuromorphic learning systems. NeuroMorse converts the top 50 words in the English language into temporal Morse code spike sequences. Despite using only two input spike channels for Morse dots and dashes, complex information is encoded through temporal patterns in the data. The proposed benchmark contains feature hierarchy at multiple temporal scales that test the capacity of neuromorphic algorithms to decompose input patterns into spatial and temporal hierarchies. We demonstrate that our training set is challenging to categorise using a linear classifier and that identifying keywords in the test set is difficult using conventional methods. The NeuroMorse dataset is available at Zenodo, with our accompanying code on GitHub at https://github.com/Ben-E-Walters/NeuroMorse.

💡 Research Summary

The paper “NeuroMorse: A Temporally Structured Dataset For Neuromorphic Computing” addresses a critical gap in the evaluation of neuromorphic computing systems. While neuromorphic engineering aims to leverage the brain’s efficient, event-driven, and temporally precise processing, most existing benchmarks are derived from static spatial datasets (like MNIST) or event-based datasets where temporal information is not crucial for task performance. This fails to capture the unique strengths of spiking neural networks (SNNs) in processing time-encoded information.

To bridge this gap, the authors introduce NeuroMorse, a novel dataset explicitly designed to prioritize temporal structure over spatial complexity. The core idea is to encode information purely in the timing of spikes. The dataset is generated by converting the top 50 most frequent English words into temporal Morse code sequences. Morse code’s dots (·) and dashes (-) are mapped to only two input spike channels. The complexity arises from the hierarchical temporal structure: specific, fixed time intervals are inserted between consecutive dots/dashes within a character (5 timesteps), between characters within a word (10 timesteps), and between words (15 timesteps). This creates nested temporal patterns that algorithms must decompose.

The dataset consists of a training set (individual sequences for the 50 keywords) and a more challenging test set. The test set is generated from a corpus of Simple Wikipedia sentences, forming a continuous spike stream. The primary task is to detect the occurrence and location of the 50 keywords within this stream, moving beyond simple classification to temporal pattern discovery in a continuous signal. To enhance realism and robustness testing, the authors introduce three types of noise—spike dropout (omission), temporal jitter (Gaussian shift in spike timing), and additive Poisson spike trains—each at three levels (None, Low, High), resulting in 27 total dataset variations.

Preliminary experiments demonstrate the dataset’s difficulty and its departure from conventional problems. A simple linear classifier based on the membrane potential trace of a single leaky integrate-and-fire (LIF) neuron receiving the two input channels achieved only 4% accuracy (2 out of 50) on the training set, barely above chance. Furthermore, a basic unsupervised network using spike-timing-dependent plasticity (STDP) with lateral inhibition struggled to reliably identify keywords in the test stream. These results underscore that NeuroMorse requires learning capabilities that go beyond simple spatial feature extraction or rate-based coding, demanding genuine temporal sequence processing.

In summary, NeuroMorse is a purpose-built benchmark that forces neuromorphic algorithms to engage with information encoded in precise temporal hierarchies. By minimizing spatial dimensions (only 2 channels) and maximizing temporal structural complexity, it provides a valuable tool for developing and evaluating SNNs on their core promised advantage: efficient processing of time-based patterns. The dataset, available in HDF5 format on Zenodo with code on GitHub, is positioned to stimulate progress in temporal learning algorithms for neuromorphic computing.