Event-Driven Neural Computing Architecture for Low-Power STDP
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current neuromorphic and synaptronic systems lack an efficient event-driven architecture for neural networks that effectively mimics biological brain functionality, particularly in implementing spike-timing dependent plasticity (STDP) for learning rules in a low-power digital format.
Innovation Solution
A low-power event-driven neural computing architecture is developed, featuring a digital CMOS spiking circuit with a crossbar memory synapse array that interconnects electronic neurons, utilizing a controller to coordinate spike events and an encoder/decoder to ensure one-to-one correspondence with software models, enabling STDP and efficient synaptic weight updates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If traditional digital models are used for neural network computation, then computational precision is maintained, but power consumption increases significantly
Solution Approach 1:
The patent implements event-driven computation where neurons and synapses are activated only periodically when spike events occur, rather than continuous operation. This allows the system to maintain computational precision while significantly reducing power consumption by entering low-power states between events.
Solution Approach 2:
The patent replaces traditional voltage-based digital computation with a spike-timing-based event-driven model that mimics biological neural computation. This substitution enables the system to achieve both low power consumption and high computational efficiency by processing information through discrete spike events rather than continuous voltage levels.
2Use of energy by moving object
If event-driven architecture is implemented to reduce power consumption, then power efficiency improves, but implementation complexity increases
Solution Approach 1:
The patent segments the neural network into discrete functional units (neurons, synapses, axons, dendrites) that operate independently based on spike events. This segmentation allows each unit to be implemented with simple digital logic, reducing overall implementation complexity while maintaining power efficiency.
Solution Approach 2:
The patent uses digital circuits to copy and simulate biological neural components, implementing STDP learning rules through digital logic that mirrors biological spike-timing mechanisms. This copying approach simplifies implementation by using well-understood digital circuit design patterns rather than requiring complex analog or mixed-signal circuits.
3Adaptability or versatility
If STDP learning rules are implemented in digital format, then learning capability is achieved, but circuit complexity increases
Solution Approach 1:
The patent pre-computes and stores lookup tables for STDP weight updates, allowing the learning rule to be implemented through simple table lookups and additions rather than complex real-time calculations. This preliminary action reduces circuit complexity while maintaining full STDP learning capability.
Solution Approach 2:
The patent introduces intermediate variables and buffer circuits that decouple the complex STDP computation from the main neural processing path. These intermediaries allow STDP learning to occur asynchronously and independently, reducing the complexity burden on the main computational circuitry.
4Speed
If crossbar synapse array is used for parallel computation, then computational speed improves, but hardware resources increase
Solution Approach 1:
The patent uses time-division multiplexing in the crossbar array, where synapses are activated periodically based on spike events rather than all synapses operating simultaneously. This allows the same hardware resources to be reused across different time steps, achieving parallel computation speed improvements without proportionally increasing hardware resources.
Data Source
AI summary
A neural network includes an electronic synapse array of multiple digital synapses interconnecting a plurality of digital electronic neurons. Each synapse interconnects an axon of a pre-synaptic neuron with a dendrite of a post-synaptic neuron. Each neuron integrates input spikes and generates a spike event in response to the integrated input spikes exceeding a threshold. A decoder receives spike events sequentially and transmits the spike events to selected axons in the synapse array. An encoder transmits spike events corresponding to spiking neurons. A controller coordinates events from the synapse array to the neurons, and signals when neurons may compute their spike events within each time step, ensuring one-to-one correspondence with an equivalent software model. The synapse array includes an interconnecting crossbar that sequentially receives spike events from axons, wherein one axon at a time drives the crossbar, and the crossbar transmits synaptic events in parallel to multiple neurons.