Efficiently Updating Connections in AI Brains

This patent describes a method for updating the connections in artificial "spiking neuron networks" more efficiently by only making changes when needed, saving computational power.

What it covers

The patent describes a way to make artificial brains, called "spiking neuron networks," learn more efficiently. Instead of constantly updating every connection, it stores a "time history of one or more inputs" (Claim 1) to a neuron. When an update is needed, the system determines "input-dependent connection change components (IDCC)" based on when inputs occurred and the current time (Claim 1). These components are then used to adjust "learning parameters" (like connection strengths) for that neuron. This allows updates to be done "on per neuron basis, as opposed to per-connection basis" (Abstract), significantly reducing the computational work. For example, if a neuron receives several inputs in a short burst, the system can calculate the combined effect of these inputs on its connections all at once, rather than processing each connection individually at every tiny time step.

What it doesn't cover

• —Neural networks that are not specifically "spiking neuron networks" (Claim 5).
• —Update methods that always process every connection individually, rather than "on per neuron basis" (Abstract).
• —Updates that are performed strictly at regular time intervals without considering "on-demand" or "event-dependent" triggers (Abstract, Claim 1).
• —Learning parameter adjustments not based on a "time history of one or more inputs" or "input-dependent connection change components" (Claim 1).

The clever bit

The core innovation is decomposing connection updates into "event-dependent connection change components" and then applying these changes "on per neuron basis" rather than checking and updating every single connection individually. This significantly reduces the number of calculations needed, especially in large networks.

Why it matters

Efficient updates are crucial for deploying artificial intelligence on devices with limited power, like smartphones or IoT sensors. By reducing the computational load for learning, this technology helps make AI more practical for "edge computing" where data is processed locally instead of in the cloud. This allows for faster responses and greater privacy in AI applications.

Real-world examples

• 1.Neuromorphic computing chips like Intel Loihi
• 2.AI accelerators for edge devices
• 3.Low-power AI systems for IoT
• 4.Spiking neural network software frameworks