Spiking Neuron Circuit Thresholding to Reduce Membrane Potential Loss
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional spiking neural networks implemented on semiconductor chips result in information loss due to the resetting of membrane potentials after voltage spike firing, leading to inaccurate outputs when performing arithmetic operations compared to digital operation circuits.
Innovation Solution
A neural network device configured with synapse and neuron circuits that utilize synaptic weights and threshold potentials to control the output of spike signals, allowing for accurate arithmetic operations with reduced energy consumption by implementing a spike-type neural network on a semiconductor device.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If conventional spiking neural networks are implemented on semiconductor chips, then energy consumption is reduced and circuit scale is minimized, but information loss occurs due to membrane potential resetting after voltage spike firing
Solution Approach 1:
The neuron circuit is divided into multiple independent neuron units, each capable of performing spike generation and membrane potential management independently. This segmentation allows parallel processing while maintaining accurate arithmetic operations in each unit, reducing overall information loss in the network.
Solution Approach 2:
The membrane potential is reset to a predetermined value (not zero) before the refractory period ends, preparing the neuron for the next spike generation. This preliminary action prevents information loss by ensuring the membrane potential starts from a known state that preserves arithmetic accuracy for subsequent operations.
Solution Approach 3:
The invention changes the parameter of membrane potential reset value from the conventional zero to a predetermined non-zero value. This parameter change maintains the energy efficiency of spiking neural networks while eliminating information loss during the reset process, enabling accurate arithmetic operations.
2Ease of operation
If membrane potentials are reset to zero after voltage spike firing, then the neuron is prepared for next firing, but arithmetic operation accuracy deteriorates due to information loss
Solution Approach 1:
The membrane potential is reset to a predetermined value in advance of the refractory period completion, ensuring the neuron is properly prepared for next firing while maintaining arithmetic accuracy. This preliminary reset action eliminates the information loss that would occur with zero resetting.
Solution Approach 2:
The system implements feedback control by monitoring the membrane potential and resetting it to a predetermined value based on the firing state. This feedback mechanism ensures that the membrane potential is always in the correct state for accurate arithmetic operations while maintaining ease of neuron firing preparation.
3Use of energy by moving object
If spiking neural networks are used instead of digital operation circuits, then energy consumption is reduced, but computational precision deteriorates due to information loss in membrane potential resetting
Solution Approach 1:
The invention changes the critical parameter of membrane potential reset value from zero to a predetermined non-zero value. This single parameter change preserves the energy efficiency of spiking neural networks while eliminating computational precision deterioration, enabling accurate arithmetic operations in the neural network device.
Solution Approach 2:
The neural network device is segmented into multiple independent neuron units with consistent reset behavior. This segmentation ensures that each unit maintains computational precision independently, and the collective network achieves both energy efficiency and accurate arithmetic operations for image recognition and classification tasks.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
The neural network device achieves accurate arithmetic operations, such as image recognition and classification, with reduced energy consumption and a small-scale circuit, without the need for CPUs or GPUs, by minimizing information loss through controlled membrane potential management.
Implementation Method 1
a charge accumulation circuit configured to accumulate charge corresponding to the synaptic current received via the first terminal and generate a membrane potential corresponding to the accumulated charge
Data Source
AI summary
A neural network device according to an embodiment includes a plurality of synapse circuits and a plurality of neuron circuits. Each of the synapse circuits acquires one or more spike signals output from one of the neuron circuits, and, in response to acquiring the spike signals, outputs a synaptic current with a current amount corresponding to an assigned synaptic weight and the spike signals. A first neuron circuit out of the neuron circuits outputs N spike signals as the one or more spike signals. The first neuron circuit includes a spike output circuit to output at least an n-th spike signal out of the N spike signals when the membrane potential is higher than an n-th threshold potential out of the N threshold potentials different from each other.