Spiking Neuron Step Function Model for Low Power Neural Network
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network models, such as those using non-leaky integrate-and-fire models with non-linear functions like exp(-x/τ), are complex and consume more power due to their intricate calculations and data movement, which complicates hardware implementation and increases power consumption.
Innovation Solution
A neural network device employing a non-leaky integrate-and-fire spiking neuron model with a postsynaptic current represented using a step function, where each neuron fires at most once per process, simplifying the model and reducing power consumption by using a step function instead of exponential functions and minimizing data movement.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If non-leaky integrate-and-fire models with non-linear functions (exp(-x/τ)) are used, then the neural network model can maintain biological realism and processing capability, but the model complexity and power consumption increase due to intricate calculations and data movement
Solution Approach 1:
The patent extracts and removes the non-linear exponential function (exp(-x/τ)) from the neural network model, replacing it with a simple step function. This extraction eliminates the computational complexity while preserving the essential spiking neuron functionality, directly resolving the contradiction between processing capability and model complexity.
Solution Approach 2:
The patent changes the mathematical parameter representation from continuous exponential decay to discrete step function transitions. By changing the functional form from exp(-x/τ) to step-based updates, the model achieves the same biological realism with dramatically reduced computational requirements, resolving the complexity contradiction.
2Reliability
If non-leaky integrate-and-fire models with non-linear functions are used, then the neural network maintains processing capability, but power consumption increases due to intricate calculations and data movement
Solution Approach 1:
The patent extracts and removes the computationally expensive exponential function calculations from the model. By eliminating these intricate calculations, the power consumption associated with complex mathematical operations is dramatically reduced while the neural network retains its essential processing capability through the step function alternative.
Solution Approach 2:
The patent replaces expensive, complex exponential calculations with cheap, simple step function operations. The step function requires minimal computational resources and can be implemented with simple logic circuits, making the neural network much more energy-efficient while maintaining functionality.
3Reliability
If complex neural network models with exponential functions are used, then processing capability is maintained, but hardware implementation becomes complicated and requires larger circuit area
Solution Approach 1:
The patent extracts the exponential function component from the model, leaving only the essential linear integration and threshold firing mechanisms. This extraction simplifies the hardware implementation by removing the need for complex exponential calculation circuits, making the system easier to manufacture while preserving core processing capability.
Solution Approach 2:
The patent substitutes complex mathematical computation (exponential functions) with simpler logical operations (step function comparisons). This substitution replaces what would require complex analog or digital computation circuits with simple logic gates and comparators, greatly easing hardware implementation.
4Reliability
If frequency-based methods are used for information transmission, then information can be transmitted reliably, but power consumption increases compared to time-based methods
Solution Approach 1:
The patent changes the information transmission parameter from frequency (number of spikes) to timing (when spikes occur). By using the temporal precision of spike arrival times rather than counting multiple spikes, the system achieves reliable information transmission with fewer spikes, thereby reducing power consumption while maintaining reliability.
Data Source
AI summary
A neural network device includes: a neuron model unit configured as a non-leaky integrate-and-fire spiking neuron and a spiking neuron with which a postsynaptic current is represented using a step function, the neuron model unit being fired once at most in one process of a neural network to indicate an output of the neural model unit itself at firing timing; and a transfer processing unit that transfers information between the neuron model unit.