Neural Network Device Stochastic Synaptic Weight Update

The patent changes the representation parameter of synaptic weights from continuous values to discrete binary values (0 or 1). This parameter change fundamentally transforms the memory requirements and computational approach, enabling accurate learning in spiking neural networks while dramatically reducing memory capacity needs and device complexity.

The patent employs stochastic update rules that probabilistically modify synaptic weights based on spike timing relationships. Instead of requiring precise continuous weight storage, the system uses simple binary weights that are updated stochastically during learning, effectively replacing complex memory systems with simpler, more efficient components.

The patent replaces the mechanical/analog memory system with a digital/stochastic update system. Instead of using analog memory components that require precise signal control and complex circuitry, the system uses binary weight representation with stochastic update rules implemented through simpler digital circuits and probability control mechanisms.

The patent fundamentally changes the storage parameter from continuous analog values to discrete binary values. This parameter transformation eliminates the need for complex analog memory circuits while maintaining learning accuracy through the stochastic update mechanism that captures essential temporal relationships between spikes.

The patent employs stochastic update rules that enable discrete weight systems to automatically adapt and refine their representations during learning. The probability control circuit and update mechanism allow the system to self-adjust synaptic weights based on spike timing relationships, compensating for the limitations of discrete representation and achieving precise learning outcomes.

The patent changes the update mechanism parameter from deterministic to stochastic. This allows discrete binary weights to evolve through probabilistic updates driven by temporal relationships between pre- and post-synaptic spikes, enabling the system to capture precise temporal information despite the discrete nature of the weights.

The patent incorporates feedback mechanisms through the probability control circuit that monitors spike timing relationships and adjusts update probabilities accordingly. This feedback loop ensures that stochastic updates accurately reflect the temporal correlations between pre- and post-synaptic spikes, maintaining update accuracy while using simpler circuits and lower power.

The stochastic update mechanism serves itself by automatically adjusting weight values based on observed spike timing patterns. The system self-regulates the update process through probability control, ensuring accurate learning without requiring complex external control circuits or additional computational resources.