One-core neuron deep learning for time series prediction.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Large language models and large vision models face challenges due to high computational demands and resource consumption.
• The need for efficient and interpretable deep learning frameworks is critical.

Purpose of the Study:

• To propose a novel 'small model' framework, the one-core-neuron system (OCNS), to address the limitations of large-scale models.
• To demonstrate that OCNS can achieve performance comparable to large models with significantly fewer parameters.

Main Methods:

• The OCNS framework utilizes a single core neuron with multiple delay feedback.
• This design enables the conversion of input feature vectors into one-dimensional time-series, theoretically capturing system dynamics.
• Spatiotemporal information transformation is leveraged for forecasting tasks.

Main Results:

• The OCNS framework significantly reduces model parameters while maintaining comparable performance to large models.
• The system exhibits excellent and robust performance in time-series forecasting, particularly for short-term, high-dimensional systems.
• The interpretability of the single-core-neuron design is highlighted.

Conclusions:

• The proposed OCNS offers a new paradigm for constructing efficient deep learning frameworks using small models.
• OCNS presents substantial potential for achieving efficient deep learning with reduced computational requirements.
• The framework provides insights into developing interpretable and resource-efficient AI systems.