Learning Universal Computations with Spikes

Area of Science:

• Computational neuroscience
• Artificial intelligence

Background:

• Spiking neural networks (SNNs) are crucial for understanding information processing in animals.
• Complex computations, like delayed reactions and self-sustained activity, are essential for SNNs.
• Many SNN computations necessitate the internal modeling of the external world.

Purpose of the Study:

• To demonstrate how SNNs can perform complex computations and generate internal world models.
• To identify constraints enabling SNNs as substrates for general-purpose computing.
• To explore learning rules for SNNs to tackle challenging benchmark tasks.

Main Methods:

• Deriving constraints for SNNs, including dendritic/synaptic nonlinearities and constrained connectivity.
• Combining SNNs with learning rules for output and recurrent connections.
• Applying these SNNs to benchmark tasks like chaotic dynamics generation and memory-dependent computations.

Main Results:

• Identified network constraints that enable powerful, general-purpose computation in SNNs.
• Demonstrated successful learning of complex tasks, including self-sustained chaotic dynamics.
• Showcased SNNs' ability to build world models and utilize them for system control.

Conclusions:

• Constrained SNNs with nonlinearities and learning rules can perform sophisticated computations.
• These networks can generate internal world models for advanced cognitive functions.
• The findings provide a neurobiological basis for complex information processing and control in SNNs.