Memristive Neuromorphic Circuit Training via Logarithmic Voltage
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current memristive neuromorphic circuits face challenges in training complex neural networks due to non-linear kinetics of memristive devices, leading to inefficiencies in hardware implementation, particularly with methods like ex-situ training and time-consuming closed-loop tuning, which may not scale for demanding applications.
Innovation Solution
A method for training memristive neuromorphic circuits that involves sensing input and error voltages, computing desired conductance changes, and applying training voltages proportional to the logarithmic value of these changes, allowing for efficient conductance modification without complex feedback tuning circuitry, enabling scalable training of neural networks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If closed-loop tuning control is used for in-situ training, then training accuracy is improved, but training time and circuit complexity increase significantly
Solution Approach 1:
The patent extracts the complex feedback control mechanism from the training process and replaces it with a direct voltage application method. By removing the need for continuous monitoring and adjustment loops, the system achieves training without the time-consuming feedback iterations while maintaining acceptable accuracy through the logarithmic voltage relationship.
Solution Approach 2:
The patent applies preliminary action by directly calculating the required conductance change before training and applying the corresponding logarithmic voltage in a single step. This eliminates the need for iterative feedback adjustments, as the correct training voltage is determined and applied in advance based on the desired weight update.
2Measurement precision
If closed-loop tuning control is implemented, then training accuracy is improved, but device complexity increases
Solution Approach 1:
The patent removes the complex feedback control circuitry from the system by replacing it with a direct voltage application approach. The feedback loop, sensors, and iterative adjustment mechanisms are extracted and eliminated, leaving a simpler circuit that applies pre-calculated logarithmic voltages directly to the memristive devices.
3Productivity
If ex-situ training is used, then training speed is improved, but hardware variability is not accounted for
Solution Approach 1:
The patent combines preliminary action with hardware awareness by pre-calculating the logarithmic training voltages based on the actual hardware characteristics and variability. This allows the system to prepare accurate training voltages in advance that account for device-specific variations, achieving both speed and reliability.
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 efficient training of memristive neuromorphic circuits, improving performance and scalability by simplifying the training process and reducing the need for complex feedback mechanisms, thus addressing the limitations of existing methods.
Implementation Method 1
Memristive devices are well-suited for implementing synaptic weights in neural networks due to their resistive switching characteristics being able to model the synaptic weights in analog form
Implementation Method 2
applying a training voltage to the memristive device, the training voltage being proportional to a logarithmic value of the desired conductance change
Data Source
AI summary
A neural network is implemented as a memristive neuromorphic circuit that includes a neuron circuit and a memristive device connected to the neuron circuit. An input voltage is sensed at a first terminal of a memristive device during a feedforward operation of the neural network. An error voltage is sensed at a second terminal of the memristive device during an error backpropagation operation of the neural network. In accordance with a training rule, a desired conductance change for the memristive device is computed based on the sensed input voltage and the sensed error voltage. Then a training voltage is applied to the memristive device. Here, the training voltage is proportional to a logarithmic value of the desired conductance change.