How Artificial Neurons Learn from Spikes and Internal States

This patent describes a method for artificial neurons to learn by adjusting their connections based on a history of incoming electrical 'spikes' and the neuron's own internal state, aiming to improve its performance.

What it covers

This patent outlines a computer-implemented method for learning in artificial spiking neuron networks. The core idea is to update the strength of connections (called 'synaptic weights') between neurons. It does this by tracking 'traces' for each connection, which are like a memory of past inputs (Claim 1). When a neuron receives a 'spiking input' (like a brief electrical signal), it calculates a 'rate of change' for these traces. This rate depends on the trace's previous value and a unique combination of a 'neuron portion' (reflecting the neuron's overall state) and a 'connection portion' (reflecting the specific input's history). For example, a robot's control system using these neurons could learn to associate specific sensor inputs (spikes) with desired motor outputs by adjusting connection strengths based on how well the robot is performing a task, with the 'node state' transitioning towards a 'target state' (Claim 2).

What it doesn't cover

• —Does not cover learning methods for traditional artificial neural networks that do not use 'spiking input' or 'spiking neurons' (Claim 1).
• —Does not cover learning rules that update connection strengths without using 'traces' that store a time-history of inputs (Claim 1).
• —Does not cover learning where the update of connection strengths is not dependent on a 'product of a neuron portion and a connection portion' (Claim 1).
• —Does not cover learning where the 'node component' (neuron's state) is not common to multiple connections or interfaces (Claim 2).
• —Does not cover learning where connection updates are purely continuous and not based on 'event-dependent connection change components' (Abstract).

The clever bit

The novelty lies in its 'generalized state-dependent learning framework,' which updates connection strengths by combining a 'per-neuron contribution' (based on the neuron's overall state) with a 'per-connection contribution' (based on specific input history). This allows for 'event-dependent connection changes' that can be executed on a 'per neuron basis,' making learning more efficient and biologically plausible.

Why it matters

This patent provides a framework for how artificial neurons, particularly 'spiking neurons,' can learn and adapt. Spiking neural networks are a promising area for developing more energy-efficient and brain-like artificial intelligence. The assignee, Brain Corp, is active in robotics, where such learning mechanisms are crucial for autonomous systems to adapt to new environments and tasks.

Real-world examples

• 1.Robotics control systems for autonomous vehicles
• 2.Neuromorphic computing hardware
• 3.Energy-efficient AI processors
• 4.Event-based vision processing systems