Bipartite Memristive Networks for Threshold-Aware Deep Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing memristive neural networks face challenges in efficiently training and optimizing conductances to minimize error functions due to limitations in voltage control and conductance changes, particularly with memristors having threshold voltage drops.
Innovation Solution
A method is introduced to train a bipartite memristive network using a modified backpropagation algorithm that applies threshold voltages or currents proportional to error deltas to adjust memristor conductances, ensuring conductance changes occur only within the memristive shells, thereby mimicking synaptic weight adjustments.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If threshold voltage is applied to memristor to change conductance, then conductance modification is achieved, but voltage control precision is reduced due to threshold limitation
Solution Approach 1:
The voltage control is segmented into two distinct phases: a threshold voltage phase that enables conductance change, and a sub-threshold voltage phase that provides fine-grained precision control. This segmentation allows the system to overcome the threshold limitation by separating the activation function from the precision adjustment function.
Solution Approach 2:
The patent implements dynamic voltage control by adjusting the amplitude and duration of voltage pulses applied to the memristor. By dynamically varying these parameters, the system can precisely control conductance changes while accounting for the threshold effect, enabling both coarse and fine adjustments.
2Adaptability or versatility
If backpropagation algorithm is implemented in memristive network, then learning capability is improved, but training complexity increases due to error propagation requirements
Solution Approach 1:
The patent introduces an intermediary training processor that acts as a mediator between the input/output electrodes and the memristor array. This intermediary device handles the complex error propagation calculations and voltage application logic, simplifying the overall system architecture while enabling full backpropagation functionality.
Solution Approach 2:
The patent replaces traditional electronic voltage control with a nanofiber-based mechanical system for implementing backpropagation. The nanofibers provide physical isolation and enable independent voltage application to each memristor, substituting complex electronic control circuits with a more manageable physical architecture.
3Manufacturing precision
If voltage is applied to change memristor conductance, then weight adjustment is achieved, but unintended conductance changes may occur outside memristive shell
Solution Approach 1:
The patent employs a nested structure where the memristive shell is contained within an insulating nanofiber matrix. This nesting provides physical confinement that prevents voltage-induced conductance changes from spreading to adjacent memristors, isolating each device's electrical influence to its own shell region.
Solution Approach 2:
The patent utilizes thin-film nanofiber structures with insulating shells that wrap around each memristor. These flexible thin films provide electrical isolation while maintaining close proximity between devices, preventing harmful lateral conductance changes and enabling dense packing without cross-talk.
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 effective training of memristive networks by selectively modifying conductances based on error deltas, enhancing the network's learning capabilities and output accuracy.
Implementation Method 1
A memristor is a passive non-linear two-terminal electrical component. In a memristor, the electrical resistance of the device depends on a history of current that previously flowed through it (or voltage applied across it).
Implementation Method 2
The strength of the weight can be modified over time by controlling the voltage drop across the memristor. Memristors typically have a threshold voltage drop, below which no change in conductance will occur.
Implementation Method 3
applying a threshold voltage or current to the output node for a time period proportional to a magnitude of the error delta
Data Source
AI summary
A bipartite memristive network and method of teaching such a network is described herein. In one example case, the memristive network can include a number of nanofibers, wherein each nanofiber comprises a metallic core and a memristive shell. The memristive network can also include a number of electrodes deposited upon the nanofibers. A first set of the number of electrodes can include input electrodes in the memristive network, and a second set of the number of electrodes can include output electrodes in the memristive network. The memristive network can be embodied as a bipartite memristive network and trained according to the method of teaching described herein.