Neuromorphic Crossbar Arrays for Memory-Augmented Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Regular neural network systems face limitations in information retention, as stored data can be overwritten by new inputs and is only preserved for a finite time, lacking an efficient mechanism for long-term information retention.
Innovation Solution
A memory-augmented neural network system utilizing a neuromorphic memory device with a crossbar array structure, where electronic devices are programmed to incrementally change states through write signals and vectors, allowing for efficient read and write operations without the need for partial memory resets, leveraging memristive devices like phase-change memory cells.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If regular neural network systems store data in internal states, then the system can process information, but the stored data can be overwritten by new inputs and is only preserved for a finite time
Solution Approach 1:
The patent transitions from internal state storage to external memory storage, adding a spatial dimension to information retention. The external memory unit stores data in a separate memory unit that can be accessed by the neural network, effectively moving data from transient internal states to persistent external storage, thereby resolving the contradiction between processing capability and information retention duration
Solution Approach 2:
The patent introduces an external memory unit as an intermediary between the neural network controller and data storage. This memory unit acts as a mediator that holds information indefinitely and allows the controller to retrieve and use stored data, solving the problem of finite data preservation time while maintaining neural network processing functionality
2Speed
If memory-augmented neural networks use crossbar structures for in-memory computation, then memory access speed increases, but the system complexity increases
Solution Approach 1:
The patent segments the neural network system into distinct functional units: a controller for executing neural network operations and a separate external memory unit for data storage. The memory unit itself is segmented into memory cells arranged in a crossbar structure with word lines and bit lines. This segmentation allows the system to achieve fast in-memory computation while managing complexity through modular design
Solution Approach 2:
The crossbar memory structure serves multiple functions: it acts as both the storage medium and the computational engine. The same memory cells that store data also perform multiply-accumulate operations during read operations, eliminating the need for separate processing units and reducing overall system complexity despite the advanced memory architecture
3Use of energy by moving object
If electronic devices in crossbar arrays incrementally change states through write signals, then energy consumption decreases, but the manufacturing precision requirements increase
Solution Approach 1:
The patent employs periodic write signals to incrementally change the conductance states of electronic devices in the crossbar array. Instead of applying large voltage pulses that consume significant energy, the system uses repeated small voltage increments over multiple cycles to achieve the desired state changes. This periodic action reduces energy consumption per operation while requiring precise control and timing, thereby increasing manufacturing precision requirements
Solution Approach 2:
The patent changes the operational parameters of electronic devices from binary switching to incremental conductance modulation. By programming devices to change their conductance states incrementally through controlled voltage applications, the system achieves lower energy consumption per write operation. However, this requires higher precision in manufacturing to ensure consistent and predictable incremental changes across all devices in the array
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 memory-augmented neural networks to efficiently utilize crossbar structures for in-memory computation, speeding up memory accesses and reducing energy consumption while maintaining the ability to retain information over long periods.
Implementation Method 1
leveraging memristive devices like phase-change memory cells
Implementation Method 2
Data is retrieved from the neuromorphic memory device, according to a multiply-accumulate operation, by coupling read signals into one or more of the input lines
Data Source
AI summary
In a hardware-implemented approach for operating a neural network system, a neural network system is provided comprising a controller, a memory, and an interface connecting the controller to the memory, where the controller comprises a processing unit configured to execute a neural network and the memory comprises a neuromorphic memory device with a crossbar array structure that includes input lines and output lines interconnected at junctions via electronic devices. The electronic devices of the neuromorphic memory device are programmed to incrementally change states by coupling write signals into the input lines based on: write instructions received from the controller and write vectors generated by the interface. Data is retrieved from the neuromorphic memory device, according to a multiply-accumulate operation, by coupling read signals into one or more of the input lines of the neuromorphic memory device based on: read instructions from the controller and read vectors generated by the interface.