N-Omniglot, a large-scale neuromorphic dataset for spatio-temporal sparse few-shot learning

Area of Science:

• Artificial Intelligence
• Computational Neuroscience

Background:

• Few-shot learning is a key human cognitive ability, yet challenging for current AI and spiking neural networks (SNNs).
• Existing few-shot datasets lack temporal information crucial for SNNs, hindering their development in this domain.

Purpose of the Study:

• To introduce N-Omniglot, the first neuromorphic dataset specifically designed for few-shot learning with SNNs.
• To provide a benchmark dataset that captures temporal dynamics for SNN few-shot learning research.

Main Methods:

• Developed N-Omniglot dataset using Dynamic Vision Sensor data.
• Collected 1,623 categories of handwritten characters with 20 samples per class.
• Adapted nearest neighbor, convolutional network, SiameseNet, and meta-learning algorithms into spiking versions.

Main Results:

• N-Omniglot offers high temporal coherence and spareness, addressing limitations of traditional datasets.
• The dataset facilitates the development of SNN algorithms for few-shot learning tasks.
• Spiking versions of standard algorithms were implemented for verification and benchmarking.

Conclusions:

• N-Omniglot serves as a crucial resource for advancing few-shot learning in SNNs.
• The dataset's temporal information provides a unique advantage for SNN research.
• This work establishes a new benchmark for evaluating SNNs in few-shot learning scenarios.