Delta-Sigma Neuron Circuits for Variability-Tolerant Memristive Training

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing deep neural networks face challenges in handling computationally intensive arithmetic operations due to excessive data movement between memory elements and processing units, particularly in custom hardware, and memristor technology is hindered by variability and integration issues.

Innovation Solution

Implementing delta-sigma modulators as neuron circuits in memristive synapses to encode information using spike frequency and timing, reducing data movement and enhancing training accuracy through supervised and unsupervised learning algorithms.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Use of energy by moving object

If memristors are used as synapses in deep neural networks, then energy efficiency and non-volatility are improved, but intrinsic variability and integration challenges worsen

Engineering Contradiction:

Improveenergy efficiencyVSAvoidintrinsic variability

Core Design Contradiction:

Use of energy by moving objectVSReliability

Solution Approach 1:

The patent implements a feedback mechanism where the output of each neuron is fed back to its own delta-sigma modulator through a feedback path. This feedback allows the system to continuously adjust and compensate for variations in memristor conductance values, effectively correcting for intrinsic variability while maintaining the energy-efficient non-volatile storage properties of memristors

Inventive Principle:

Principle #23Feedback

Solution Approach 2:

The patent transforms the static conductance values of memristors into dynamic, adjustable parameters through the delta-sigma modulator feedback system. By continuously adjusting the effective synaptic weights based on feedback signals and training data, the system adapts the memristor parameters to overcome manufacturing variability and achieve reliable neural network functionality

Inventive Principle:

Principle #35Parameter changes

2Extent of automation

If conventional training methods are used with memristive synapses, then training capability is achieved, but excessive data movement between memory and processing units increases energy consumption

Engineering Contradiction:

Improvetraining capabilityVSAvoidenergy consumption

Core Design Contradiction:

Extent of automationVSUse of energy by moving object

Solution Approach 1:

The patent merges the memory function (memristor synapses) with the processing function (delta-sigma modulator neurons) into a unified neuromorphic architecture. This integration eliminates the need for separate memory and processing units, thereby removing the energy-consuming data movement between them while maintaining full training capability through in-memory computation

Inventive Principle:

Principle #5Merging (Combining)

3Measurement precision

If high precision training is pursued, then accuracy is improved, but computational complexity and data movement requirements increase

Engineering Contradiction:

Improvetraining accuracyVSAvoidcomputational complexity

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The patent replaces conventional digital computation mechanisms with analog neuromorphic mechanisms. The delta-sigma modulator neurons perform computational operations using analog voltage signals and feedback loops rather than digital arithmetic, achieving high-precision training through continuous analog adjustment while reducing computational complexity and eliminating the need for extensive data movement between digital memory and processing units

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

Data Source

PatentUS12632738B2Delta-sigma modulation neurons for high-precision training of memristive synapses in deep neural networks

Publication Date: 2026.05.19 TECHNION RES & DEV FOUND LTD

AI summary

A neural network comprising: a plurality of interconnected neural network elements, each comprising: a neuron circuit comprising a delta-sigma modulator, and at least one synapse device comprising a memristor connected to an output of said neuron circuit; wherein an adjustable synaptic weighting of said at least one synapse device is set based on said output of said neuron circuit.