Online Spatio-Temporal Learning in Deep Neural Networks

Area of Science:

• Computational Neuroscience
• Machine Learning
• Artificial Intelligence

Background:

• Biological neural networks exhibit continuous online learning, unlike traditional deep learning methods.
• Error backpropagation through time (BPTT) requires offline gradient computation, limiting real-time adaptation.
• Existing methods struggle to achieve efficient online training for complex neural network architectures like SNNs.

Purpose of the Study:

• To develop a novel online learning algorithm for deep recurrent neural networks (RNNs) and spiking neural networks (SNNs).
• To enable biologically plausible, continuous adaptation in artificial neural networks.
• To achieve BPTT-equivalent gradient performance in an online setting for SNNs.

Main Methods:

• Introduced online spatio-temporal learning (OSTL), a framework inspired by biological neural processing.
• OSTL separates spatial and temporal gradient computations for efficient online processing.
• Extended OSTL to accommodate various architectures, including LSTMs and GRUs.

Main Results:

• OSTL achieves gradient equivalence to BPTT for shallow SNNs, enabling online training with comparable performance.
• Demonstrated online training of SNNs at low time complexity.
• Achieved results on par with BPTT baselines in language modeling and speech recognition tasks.

Conclusions:

• OSTL provides a viable and efficient framework for online learning in deep RNNs and SNNs.
• The proposed method bridges the gap between biological online learning and artificial neural network training.
• OSTL offers a promising direction for developing more adaptive and biologically realistic AI systems.