Benchmarking Spike-Based Visual Recognition: A Dataset and Evaluation.

Area of Science:

• Neuroscience and Artificial Intelligence
• Spike-based neural computation
• Biologically-inspired computing

Background:

• The primate visual pathway is a key area of study for understanding the brain and developing biologically-inspired computation.
• Spiking Neural Networks (SNNs) show promise in visual recognition tasks and can be implemented on neuromorphic hardware for real-time processing.
• Neuromorphic sensors like silicon retinas provide real-time visual input for mobile systems.

Purpose of the Study:

• To address the need for standardized vision benchmarks in spike-based neural processing.
• To propose a large dataset of spike-based visual stimuli and a corresponding evaluation methodology for SNNs.
• To facilitate quantitative comparisons between different SNN models and hardware implementations.

Main Methods:

• Introduction of an initial Neuromorphic Engineering (NE) dataset using MNIST digits, compatible with current spike-based image recognition research.
• Generation of spike trains using rate-based Poisson, rank order encoding, and silicon retina recordings.
• Presentation of a complementary evaluation methodology for assessing both model-level and hardware-level performance.

Main Results:

• Demonstration of the dataset and evaluation methodology using two SNN models.
• Validation of the performance of SNN models and their hardware implementations.
• Establishment of a baseline for quantitative comparison within the field.

Conclusions:

• The proposed NE dataset and evaluation methodology will promote meaningful comparisons between SNN algorithms and conventional methods.
• This work provides an assessment of the state-of-the-art in spike-based visual recognition.
• The dataset and methodology aim to guide future research directions in neural computation and advance the field.