Deep Learning with Dynamic Spiking Neurons and Fixed Feedback Weights.

Area of Science:

• Computational neuroscience
• Artificial intelligence
• Deep learning

Background:

• Deep learning excels using backpropagation in artificial neural networks.
• Biological neurons differ from artificial ones in spiking outputs, dynamic relations, and synapse coordination.
• Current deep learning algorithms are unlikely to directly operate in the brain.

Purpose of the Study:

• To investigate if deep learning algorithms can be adapted to function within the brain's biological constraints.
• To develop novel learning rules and algorithms for deep spiking neural networks.
• To demonstrate effective learning in deep spiking networks comparable to traditional deep learning.

Main Methods:

• Developed learning rules to approximate dynamic input-output relations with piecewise-smooth functions.
• Adapted the feedback alignment algorithm to train deep networks without coordinated forward and feedback synapses.
• Implemented an intracellular teaching signal reflecting neuronal nonlinearity in spiking networks.

Main Results:

• Demonstrated that deep spiking networks can learn effectively despite biological differences from artificial networks.
• Showed useful learning in synapses up to nine layers upstream from output cells.
• Achieved competitive performance on the MNIST digit recognition task compared to existing spiking network models.

Conclusions:

• The proposed methods enable deep learning algorithms to operate effectively in spiking neural networks, bridging the gap between AI and neuroscience.
• These findings suggest a viable computational framework for brain-inspired AI.
• The developed techniques offer a pathway for more biologically plausible and potentially more efficient deep learning models.