Variable Sampling Frequency for Spiking Neural Network Power Reduction

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing digital spiking neural networks face challenges in reducing power consumption while maintaining accuracy and stability, as most power-conscious implementations are limited by their hardware design and are not adaptable to varying biological detail, leading to inefficiencies in resource usage and noise susceptibility.

Innovation Solution

The Input-Dependent Variable Sampling (I-DEVS) method modulates the sampling frequency of artificial neurons based on input signals, reducing unnecessary switching activity by adjusting the frequency inversely proportional to the input current, thereby reducing power consumption and hardware resource usage without affecting the neuron's behavior.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a fixed high sampling frequency is used in digital spiking neurons, then accuracy and stability are maintained, but power consumption increases due to unnecessary switching activity during low-stimulus states

Engineering Contradiction:

Improveneuron output accuracyVSAvoiddynamic power consumption

Core Design Contradiction:

Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The patent applies dynamics by making the sampling frequency variable rather than fixed. The sampling frequency dynamically adjusts based on the input stimulus level - using higher frequencies during high-stimulus states to maintain accuracy, and lower frequencies during low-stimulus states to reduce power consumption. This is implemented through a state machine that monitors neuron state and adjusts the sampling rate accordingly, resolving the contradiction between maintaining measurement precision and reducing energy usage.

Inventive Principle:

Principle #15Dynamics

Solution Approach 2:

The patent changes the parameter of sampling frequency from a constant value to a variable parameter that adapts to input conditions. By modifying the sampling frequency parameter based on stimulus level and neuron state, the system achieves both high accuracy when needed and low power consumption when not needed, directly addressing the technical contradiction between precision and energy efficiency.

Inventive Principle:

Principle #35Parameter changes

2Use of energy by moving object

If analog or mixed-signal VLSI systems are used for neuron implementation, then power consumption is reduced, but noise susceptibility increases and design complexity increases

Engineering Contradiction:

Improvepower consumptionVSAvoidnoise susceptibility

Core Design Contradiction:

Use of energy by moving objectVSObject-affected harmful factors

Solution Approach 1:

The patent makes the digital neuron implementation universally applicable to any neuron model while achieving power savings. The I-DEVS method with state-machine control can be applied to different neuron models (Leaky Integrate-and-Fire, AdEx, Izhikevich, etc.) without requiring analog components. This universal digital approach reduces noise susceptibility compared to analog systems while maintaining low power consumption through variable sampling frequency.

Inventive Principle:

Principle #6Universality (Multi-functionality)

3Use of energy by moving object

If CORDIC algorithm is used for power reduction in digital neurons, then hardware resources and power are reduced, but system throughput decreases due to iterative computation

Engineering Contradiction:

Improvepower consumptionVSAvoidsystem throughput

Core Design Contradiction:

Use of energy by moving objectVSProductivity

Solution Approach 1:

The patent applies periodic action by using event-driven computation where the neuron is only evaluated at specific sampling intervals rather than continuously. The state machine controls periodic updates based on input events and neuron state transitions. This periodic evaluation approach reduces power consumption by eliminating continuous computation while maintaining system throughput by updating only when necessary, avoiding the iterative overhead of CORDIC algorithms.

Inventive Principle:

Principle #19Periodic action

4Use of energy by moving object

If simple neuron models with low biological detail are used, then power consumption is reduced, but biological accuracy and behavioral replication are compromised

Engineering Contradiction:

Improvepower consumptionVSAvoidbiological behavior accuracy

Core Design Contradiction:

Use of energy by moving objectVSReliability

Solution Approach 1:

The patent applies dynamics to the sampling frequency based on the specific biological neuron model being implemented. Different neuron models (Leaky Integrate-and-Fire, Adaptive Exponential Integrate-and-Fire, Izhikevich) have different biological details and requirements. The state machine dynamically adjusts sampling to match the temporal characteristics of each model, enabling accurate replication of biological behavior for complex models while maintaining low power consumption by using coarser sampling when appropriate for the specific model's time constants and dynamics.

Inventive Principle:

Principle #15Dynamics

Data Source

PatentUS20230153585A1Dual State Circuit for Energy Efficient Hardware Implementation of Spiking Neural Networks

Publication Date: 2023.05.18 UNIVERSITY OF WINDSOR

AI summary

In a preferred embodiment, there is provided a method for reducing power consumption in an artificial neural network, the method comprising: receiving an input signal; modulating a sampling frequency of an artificial neuron based on the input signal; and forwarding the input signal or a further input signal obtained from the input signal to the artificial neuron at the sampling frequency, wherein said modulating the sampling frequency comprises increasing the sampling frequency with an increased input signal and reducing the sampling frequency with a decreased input signal.