Memristor Crossbar Array On-Chip Learning System
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing computers face the Von Neumann bottleneck when implementing convolutional neural networks, leading to slow and costly separation of calculation and storage, and are limited to pre-trained functions that cannot solve technical problems in a timely and flexible manner.
Innovation Solution
A convolutional neural network on-chip learning system based on non-volatile memory, which includes an input module, a convolutional neural network module, an output module, and a weight update module, utilizing memristors to simulate synaptic functions and update weights through conductance modulation, allowing for on-chip learning and efficient information processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Power
If convolutional neural networks are implemented on traditional CPU/GPU systems, then computational capabilities are achieved, but power consumption increases significantly due to the Von Neumann bottleneck
Solution Approach 1:
The patent merges storage and computation functions by implementing synaptic weights directly in memristor devices. The crossbar array structure allows weight storage in the conductance states of memristors while simultaneously performing matrix multiplication through parallel current summation, eliminating the need to move data between separate storage and processing units.
Solution Approach 2:
The patent replaces traditional electronic computation mechanics (CPU/GPU sequential processing) with a physical field-based system where computation emerges from parallel electrical current flow through the memristor crossbar array. Matrix multiplication is performed physically through Ohm's law and Kirchhoff's current law rather than through sequential electronic operations.
2Adaptability or versatility
If convolutional neural networks are implemented on traditional computer systems, then specific pre-trained functions are achieved, but the system cannot solve technical problems in a timely and flexible manner
Solution Approach 1:
The patent implements on-chip learning capability where the neural network can perform backpropagation and weight updates directly on the hardware without external computer assistance. The system uses the same crossbar array and readout circuits to perform both forward propagation and gradient computation, enabling autonomous learning and adaptation.
3Productivity
If separation of calculation and storage is implemented in traditional systems, then functional modularity is achieved, but processing speed decreases due to data transfer overhead
Solution Approach 1:
The patent merges storage and computation functions by implementing synaptic weights directly in memristor devices. The crossbar array structure allows weight storage in the conductance states of memristors while simultaneously performing matrix multiplication through parallel current summation, eliminating the need to move data between separate storage and processing units.
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
This approach enables real-time, low-energy simulation of brain-scale neural networks, overcoming the limitations of the Von Neumann system by integrating storage and calculation, improving operation speed and density, and enabling flexible on-chip learning.
Implementation Method 1
The on-chip learning implements a synaptic function by using the conductance modulation characteristic that the conductance of the memristor changes according to the applied pulse
Data Source
AI summary
Disclosed by the disclosure is a convolutional neural network on-chip learning system based on non-volatile memory, comprising: an input module, a convolutional neural network module, an output module and a weight update module. The on-chip learning of the convolutional neural network module implements the synaptic function by using the characteristic of the memristor, and the convolutional kernel value or synaptic weight value is stored in a memristor unit; the input module converts the input signal into the voltage signal; the convolutional neural network module converts the input voltage signal layer-by-layer, and transmits the result to the output module to obtain the output of the network; and the weight update module adjusts the conductance value of the memristor in the convolutional neural network module according to the result of the output module to update the network convolutional kernel value or synaptic weight value.