Learning to Recognize Actions From Limited Training Examples Using a Recurrent Spiking Neural Model.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Learning from limited data is a key challenge in machine learning.
• Spiking neural networks (SNNs) offer a biologically inspired approach to information processing.
• Action recognition in videos requires understanding complex temporal dynamics.

Purpose of the Study:

• To develop a reservoir-based spiking neural model for few-shot action recognition.
• To introduce a novel spike encoding method inspired by microsaccades for enhanced temporal feature extraction.
• To establish a new benchmark for SNNs in limited-data action recognition.

Main Methods:

• A reservoir-based spiking neural network architecture was employed.
• A novel encoding method, inspired by microsaccades, was developed to extract spike information from video frames.
• The model was trained and evaluated on the UCF-101 dataset using a few-shot learning paradigm.

Main Results:

• The proposed spiking neural model achieved 81.3% Top-1 and 87% Top-5 accuracy on the UCF-101 dataset.
• The model required only 8 video examples per class for effective training, demonstrating strong few-shot learning capabilities.
• The novel encoding method successfully preserved temporal correlations across video frames.

Conclusions:

• The developed reservoir-based spiking neural model is effective for action recognition with limited labeled video data.
• This approach sets a new benchmark for spiking neural models in few-shot action recognition.
• The model demonstrates competitive accuracy compared to state-of-the-art non-spiking neural models.