NeuroMorse: A Temporally Structured Dataset For Neuromorphic Computing

Technical Summary of NeuroMorse: A Temporally Structured Dataset For Neuromorphic Computing

Problem Statement
Neuromorphic engineering seeks to mimic the brain's efficient processing by utilizing asynchronous temporal events (spikes) rather than synchronous data frames, thereby minimizing power consumption. While Spiking Neural Networks (SNNs) are theoretically well-suited for event-based tasks that leverage both spatial and temporal domains, current benchmarking practices often fail to adequately test these temporal capabilities. Many existing datasets, such as static image conversions (e.g., MNIST) or even event-based gesture datasets, prioritize spatial hierarchies. In these benchmarks, temporal information is often secondary; performance can remain high even when temporal data is binned or removed, suggesting that current evaluations do not fully capture the unique strengths of neuromorphic systems. There is a critical gap in datasets where temporal dynamics are the primary encoding mechanism for complex information.

Methodology: The NeuroMorse Dataset
To address this gap, the authors introduce NeuroMorse, a dataset designed specifically to prioritize temporal data representation over spatial or channel-wise representations. The dataset is constructed by converting the top 50 most frequently used words in the English language into temporal Morse code spike sequences.

• Encoding Scheme: The dataset utilizes only two input spike channels: one for "dots" and one for "dashes." Complex semantic information is encoded entirely through the temporal patterns of these spikes.
• Temporal Hierarchy: The encoding introduces a multi-scale temporal hierarchy:
  • $\Delta t_1$ (5 timesteps): Interval between consecutive Morse characters (dots/dashes).
  • $\Delta t_2$ (10 timesteps): Interval between consecutive English letters.
  • $\Delta t_3$ (15 timesteps): Interval between consecutive English words.
• Dataset Composition:
  • Training Set: Contains the top 50 English words converted into spike sequences.
  • Test Set: Generated from a corpus of 50,441 Simple Wikipedia articles (punctuation removed). The primary task is to identify occurrences of the 50 training keywords within this continuous stream. The test set includes 1,768,231 keyword instances.
  • Noise Variations: To increase difficulty, the authors introduced three types of noise (spike removal/dropout, temporal jitter, and Poissonian spike trains) at low and high levels, resulting in 27 total dataset variations.
• Format: Data is stored in Hierarchical Data Format 5 (HDF5), containing spike timestamps, channel information, and labels (including start/end times for keywords in the test set).

Key Contributions

1. Temporal-Centric Benchmark: NeuroMorse is presented as the first neuromorphic dataset that explicitly prioritizes temporal data representation, reducing spatial input to a minimal two channels while relying on temporal patterns for information density.
2. Hierarchical Complexity: The dataset embeds feature hierarchies at multiple temporal scales (dot/dash, letter, word), challenging algorithms to decompose input patterns into both spatial and temporal hierarchies.
3. Comprehensive Evaluation Suite: The paper provides a structured evaluation framework including a linear classifier, an unsupervised Spike-Timing Dependent Plasticity (STDP) network, and a supervised deep SNN, all tested on clean and noisy data.

Experimental Results
The authors evaluated the difficulty of the dataset using three distinct architectural approaches, all performed without noise to establish a baseline of difficulty:

1. Linear Classifier: A simple Leaky Integrate-and-Fire (LIF) neuron connected to an Ordinary Least Squares regression module achieved only 4% accuracy on the training set (double the random baseline of 2%), indicating that the temporal patterns are not linearly separable.
2. Unsupervised STDP Network: A network utilizing weight-dependent STDP synapses and lateral inhibition achieved a classification accuracy of only 0.19% on the test set. The authors note that the vast majority of errors were due to the misclassification of null classes (non-keyword segments), highlighting the difficulty of distinguishing signal from noise in a continuous stream.
3. Supervised Deep SNN: A 3-layer network trained using spiking backpropagation with surrogate gradients achieved 66% accuracy on the training set and 4.25% on the full test set. When the test set was filtered to remove non-keywords (leaving only a random sequence of keywords), accuracy rose to 12.46%.

Significance and Claims
The paper claims that NeuroMorse effectively demonstrates the limitations of current neuromorphic architectures in learning temporal hierarchies. The results suggest that:

• Standard linear classifiers fail to decode the temporal structure.
• Unsupervised STDP mechanisms struggle significantly with keyword identification in continuous streams, particularly regarding false positives on null classes.
• Even supervised deep learning approaches struggle to generalize from the training set to the complex, continuous test set.

The authors conclude that NeuroMorse serves as a rigorous benchmark to evaluate and improve SNNs, specifically targeting the temporal processing capabilities that are central to neuromorphic computing but often neglected in current evaluations. The dataset and code are made publicly available to facilitate further research in this domain.