Memristive Modulation Device for Noise-Robust Artificial Synapses

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing artificial neural networks are sensitive to noise, leading them to 'learn' from noise patterns and degrade their performance in signal processing tasks.

Innovation Solution

Implementation of a short-term plasticity mechanism using a modulation device with a control block capable of emitting two distinct types of pulses, which modifies the equivalent conductance of memristive devices to introduce noise robustness and facilitate unsupervised learning.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If artificial neural networks use standard learning mechanisms, then they can learn patterns, but they become sensitive to noise and learn from noise patterns

Engineering Contradiction:
Improvelearning capabilityVSAvoidnoise robustness
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

The invention introduces dynamic modulation of synaptic weights through a modulation device that adjusts the impact of presynaptic impulses on postsynaptic neurons. This dynamic adjustment allows the network to adapt its learning behavior in real-time, enhancing robustness to noise while preserving pattern learning capabilities. The modulation device changes the effective weight of synapses based on temporal patterns, making the learning process more selective and less susceptible to noise.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The invention changes the parameter of synaptic weight modulation by introducing a time-dependent modulation factor. Instead of fixed weights, the system uses modulated weights that vary over time based on the modulation device's output. This parameter change enables the network to differentiate between meaningful patterns and noise by adjusting weight influence dynamically, thereby improving noise robustness while maintaining learning effectiveness.

Inventive Principle:
Principle #35Parameter changes

2Device complexity

If artificial synapses use fixed conductance values, then the network structure is simple, but the network cannot perform unsupervised learning effectively

Engineering Contradiction:
Improvesynapse structureVSAvoidunsupervised learning capability
Core Design Contradiction:
Device complexityVSAdaptability or versatility

Solution Approach 1:

The invention implements preliminary action by pre-configuring the modulation device with specific modulation characteristics before the learning process begins. The modulation device is designed with predetermined modulation patterns that prepare the synapses to respond appropriately to input patterns. This preliminary configuration enables unsupervised learning without requiring complex real-time adjustments, maintaining relatively simple synapse structure while enhancing learning capability.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The invention introduces feedback mechanisms where the modulation device receives information about network activity and adjusts synaptic modulations accordingly. This feedback loop enables the system to learn from its own activity patterns, facilitating unsupervised learning. The feedback mechanism allows the network to automatically adjust to input patterns without external supervision, improving adaptability while keeping the overall structure manageable through localized feedback loops.

Inventive Principle:
Principle #23Feedback

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

The solution enhances the robustness of artificial neural networks to noise while enabling effective unsupervised learning by integrating a short-term plasticity mechanism, improving their performance in signal processing applications.

Implementation Method 1

modifies the equivalent conductance of memristive devices (11, 12)

Methodology Applied
Scientific EffectMemristive effect:

Data Source

PatentEP3549070B1Modulation device and method, artificial synapse comprising said modulation device, short-term plasticity method in an artificial neural network comprising said artificial synapse
Publication Date: 2023.11.22 COMMISSARIAT A LENERGIE ATOMIQUE ET AUX ENERGIES ALTERNATIVES
  • EP3549070B1 patent drawingFigure 1a~2b
  • EP3549070B1 patent drawingFigure 3~4b
  • EP3549070B1 patent drawingFigure 5a~5b

AI summary

An aspect of the invention concerns a modulation device (11) comprising: - at least one memristive device (Mr1), and - a control block (Ct1), the modulation device (11) having an equivalent conductance yi(t) produced by the at least one memristive device (Mr1) and the control block (Ct1) having means to: - receive a clock signal (clk) and perform a first modification of the equivalent conductance yi(t) upon receipt of each clock signal (clk), - receive an input voltage pulse (Vin) and perform a second modification of the equivalent conductance yi(t) upon receipt of each input voltage pulse (Vin), the first and second modifications being in opposite directions.