Neural Network Circuit Using Magnetic Tunnel Junction Synapses
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network circuit devices do not achieve optimal structure and power efficiency in storing synaptic coupling weights, leading to suboptimal performance in terms of power consumption and speed.
Innovation Solution
A neural network circuit device incorporating synapse circuits with magnetic tunnel junction elements and neuron circuits featuring a floating gate and control gates, which store synaptic coupling weights in a non-volatile manner and transmit them as voltage signals, allowing for efficient weight updating and high-speed operation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If synaptic coupling weight is stored in non-volatile storage element, then power consumption for retaining memory is reduced, but device optimization and performance are insufficient
Solution Approach 1:
The patent combines multiple functions into unified circuit blocks: the synapse circuit integrates weight storage (via magnetic tunnel junction element resistance states) and weight application (via voltage signal output to neuron circuit), while the neuron circuit integrates signal integration (via floating gate capacitance) and spike generation. This functional merging eliminates separate memory and processing units, reducing overall device complexity and improving optimization while maintaining non-volatile power efficiency.
Solution Approach 2:
The magnetic tunnel junction element serves multiple purposes: it stores synaptic weights non-volatently through resistance states and simultaneously enables weight application through voltage signal generation. Similarly, the floating gate in the neuron circuit both integrates incoming voltage signals and maintains the threshold potential without requiring continuous power. This multi-functionality achieves both power efficiency and high performance.
2Ease of manufacture
If conventional neural network circuit structure is used, then implementation is straightforward, but power consumption is high and speed is limited
Solution Approach 1:
The patent replaces conventional volatile memory elements and separate memory-access mechanisms with non-volatile magnetic tunnel junction elements that inherently maintain weight states without power. The neuron circuit replaces continuous monitoring mechanisms with a floating gate that passively integrates voltage signals and holds threshold potential through capacitance, eliminating the need for active power maintenance and reducing both power consumption and circuit complexity.
3Use of energy by stationary object
If synaptic weight is stored in non-volatile memory, then power for retention is saved, but weight updating and transmission efficiency are suboptimal
Solution Approach 1:
The patent utilizes the resistance parameter of the magnetic tunnel junction element to encode synaptic weights, where different resistance states correspond to different weight values. During weight updates, the resistance parameter is changed through magnetic switching, and during transmission, the resistance parameter is converted to a voltage signal parameter that directly modulates the neuron circuit. This parameter transformation enables both non-volatile storage and high-speed updating/transmission.
Solution Approach 2:
The voltage signal generated by the synapse circuit acts as an intermediary between the non-volatile magnetic tunnel junction weight storage and the neuron circuit processing. This voltage intermediary enables rapid weight application without requiring readout of the magnetic state, achieving high-speed weight transmission while maintaining non-volatile power efficiency.
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 solution reduces power consumption and enables high-speed operation by efficiently storing and updating synaptic weights, optimizing the neural network circuit device's configuration.
Implementation Method 1
a magnetic tunnel junction element, which are connected in series
Implementation Method 2
a neuron MOS transistor having a floating gate and a plurality of control gates which are capacitively coupled to the floating gate
Implementation Method 3
a writing unit allowing the magnetic tunnel junction elements to have different magnetization states by allowing a write current based on the synaptic coupling weight to be stored to flow in each magnetic tunnel junction element
Data Source
AI summary
There is provided a neural network circuit device including a plurality of synapse circuits storing a synaptic coupling weight and a neuron circuit connected to the plurality of synapse circuits. The plurality of synapse circuits store the synaptic coupling weight in a non-volatile manner and output a voltage signal having a magnitude based on the stored synaptic coupling weight in response to an input signal. The neuron circuit includes a neuron MOS transistor having a floating gate and a plurality of control gates which are capacitively coupled to the floating gate and to which the voltage signals from the plurality of synapse circuits are input respectively, and a pulse generator outputting a pulse signal by turning on or off the neuron MOS transistor.