Volatile Conducting Bridge Memory for Neuromorphic Training Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Memristor-based crossbar arrays in neuromorphic computing face inaccuracies due to sneak path currents, thermal noise, and process variations, leading to cumulative errors in training, which existing compensation methods struggle to fully address without extensive computational resources.
Innovation Solution
The implementation of bimodal volatile/non-volatile resistive memory devices, also known as volatile conducting bridges (VCBs), which exhibit self-refreshing behavior below a threshold voltage, disregarding incidental sneak currents and noise while retaining intended programming through non-volatile behavior above the threshold.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If memristor-based crossbar arrays are used to represent synapse weight matrix, then neuromorphic computing functionality is achieved, but sneak path currents and noise cause cumulative errors in training accuracy
Solution Approach 1:
The patent segments the memory device functionality into two distinct operational modes: volatile mode for disregarding sneak currents and noise during training, and non-volatile mode for retaining programmed weights. This segmentation allows the same hardware to handle different functional requirements separately, resolving the contradiction between achieving neuromorphic functionality and maintaining training accuracy.
Solution Approach 2:
The memory device dynamically switches between volatile and non-volatile behavior based on the applied voltage threshold. During training operations, the device operates in volatile mode to reject sneak path currents and noise. During weight retention, it switches to non-volatile mode. This dynamic behavior enables the system to maintain both neuromorphic functionality and training accuracy without compromise.
2Measurement precision
If compensation techniques are applied to address inaccuracies, then some error correction is achieved, but huge computational resources are required and accuracy cannot be guaranteed
Solution Approach 1:
The patent implements preliminary action by designing memory devices with intrinsic threshold-based volatile behavior that automatically disregards sneak currents and noise before they can cause cumulative errors. This preventive approach eliminates the need for post-training compensation techniques, thereby reducing computational resource requirements while guaranteeing training accuracy.
3Duration of action of stationary object
If memory devices retain all input signals including sneak currents, then continuous memory state updates occur, but cumulative errors accumulate across the array
Solution Approach 1:
The patent applies local quality by making different parts of the memory device respond differently to input signals based on their magnitude. Signals below the threshold (including sneak currents and noise) are locally rejected through volatile behavior, while signals above the threshold (intentional programming signals) are retained through non-volatile behavior. This local differentiation resolves the contradiction between memory state retention and weight accuracy.
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 eliminates the need for compensation techniques by ensuring accurate training and efficient operation, as untargeted memory devices self-refresh and remain unchanged, avoiding cumulative errors and reducing computational requirements.
Implementation Method 1
a conductive bridge to form in the oxide layer based on interaction between the oxide layer and an electrode of the memory device in response to an input voltage
Data Source
AI summary
An example device in accordance with an aspect of the present disclosure includes an active oxide layer to form and dissipate a conductive bridge. The conductive bridge is to dissipate spontaneously within a relaxation time to enable the memory device to self-refresh according to volatile behavior in response to the input voltage being below a threshold corresponding to disregarding sneak current and noise of a crossbar array in which the memory device is to operate. The conductive bridge is to persist beyond the relaxation time to enable the memory device to retain programming for neuromorphic computing training according to non-volatile behavior of the memory device in response to the input voltage not being below the threshold.