Programmable Artificial Neuron Using Non-Volatile Resistive Memory
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing artificial neurons implemented with CMOS technology lose parameter values when power is turned off and do not allow high parameter values due to large leakage currents, necessitating a more efficient and non-volatile programming method.
Innovation Solution
The use of non-volatile resistive memory to program control parameters of artificial neurons, allowing any parameter value without resetting, and enabling independent programming of each parameter through multiplexers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If CMOS technology is used to implement artificial neurons, then the circuit can be integrated and controlled, but the parameter values are lost when power is turned off
Solution Approach 1:
The patent combines CMOS circuitry with non-volatile resistive memory elements (such as OxRAM, CBRAM, or PCM) to create a hybrid neuron implementation. The resistive memory stores parameter values (weights, thresholds, time constants) non-volitilely, while the CMOS circuit provides signal processing and control functions. This merging allows the system to retain parameter information without power while maintaining the benefits of integrated circuit manufacturing.
Solution Approach 2:
The patent pre-programs parameter values into the resistive memory elements before the neuron circuit operates. These parameter values are stored in the high-impedance state of the resistive memory and are automatically loaded into the CMOS circuit when needed, eliminating the need for repeated programming and ensuring parameter retention across power cycles.
2Adaptability or versatility
If CMOS transistors are used to control parameter values, then parameters can be adjusted dynamically, but high parameter values cannot be achieved due to large leakage currents
Solution Approach 1:
The patent changes the physical state or property of the resistive memory elements to achieve different parameter values. By controlling the resistance of the memory element (through voltage or current programming), the neuron's parameters (such as time constant or threshold) are adjusted. This allows a wide range of parameter values to be achieved without being limited by transistor leakage currents, as the resistance can be tuned over several orders of magnitude.
Solution Approach 2:
The resistive memory element acts as an intermediary between the control voltage and the neuron's functional parameters. Instead of directly controlling the neuron's behavior with control voltages that suffer from leakage issues, the control voltage first programs the resistance of the memory element, which then stabilizes and provides the desired parameter value to the neuron circuit, isolating the control mechanism from leakage effects.
3Device complexity
If parameter values are reset when power is turned off, then the circuit simplifies by not requiring non-volatile storage, but information about parameter values is lost
Solution Approach 1:
The patent merges volatile CMOS memory with non-volatile resistive memory to create a system that automatically retains parameter information. The resistive memory maintains the parameter values in its resistance state even without power, and the CMOS circuit reads these values when powered on, eliminating the need for external memory storage or complex reset prevention mechanisms while preserving information.
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 solution ensures that parameter values are retained even when power is off and allows for any parameter value to be adjusted, enhancing the efficiency and bio-realism of artificial neurons.
Implementation Method 1
The invention makes it possible to programme the control parameters of an artificial neuron in a non-volatile manner using a resistive memory
Implementation Method 2
a multiplexer Mux having a selection input Vsel and two outputs, a first output corresponding to a programming state 0 and a second output to an operating state 1, the selection input allowing the choice between the two outputs
Data Source
Figure 1~2
Figure 3~4A
Figure 4B
AI summary
One aspect of the invention relates to a programmable artificial neuron emitting an output signal controlled by at least one control parameter, characterized in that it comprises, for each control parameter, a capacitor and at least one block (200) comprising: - at least one multiplexer (Mux) configured to be in two states: a programming state (0) and an operating state (1); - a transistor (T); and - a non-volatile resistive memory (MR) connected in series with said transistor (T), said capacitor and said resistive memory (MR) being mounted in parallel; said multiplexer (Mux) being configured to: ∘ when it is in the programming state (0), fix a value of the resistance of the resistive memory (MR) to fix the value of the control parameter; ∘ when it is in the operating state (1), retain the value of the fixed resistance of the resistive memory (MR).