Bayesian Neural Network Circuit Using Jitter and Charge-Based Weights

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Traditional neural networks suffer from overfitting and sensitivity to malicious attacks due to the lack of uncertainty information, leading to performance issues and reliability concerns.

Innovation Solution

Implementing a Bayesian neural network (BNN) with a C-2C ladder, jittery oscillator, and charge pump to generate weights following a Gaussian distribution, leveraging analog computing-in-memory (CiM) for reduced hardware and energy consumption.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If traditional neural networks are used, then the system is simple to implement, but the network suffers from overfitting and sensitivity to malicious attacks due to lack of uncertainty information

Engineering Contradiction:
Improverobustness against overfitting and attacksVSAvoidnetwork structure complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent transforms fixed weight parameters into probabilistic distributions characterized by mean and variance. This parameter transformation enables the network to represent uncertainty explicitly, improving robustness against overfitting and adversarial attacks while maintaining the same fundamental network architecture.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent pre-computes and stores mean and variance values for each weight during the training phase. These pre-computed statistical parameters are then used during inference to generate sample weights, eliminating the need for complex real-time probabilistic computations and reducing operational complexity.

Inventive Principle:
Principle #10Preliminary action

2Reliability

If Bayesian neural network with probability distributions is implemented, then uncertainty information is captured to improve reliability, but hardware resources and energy consumption increase

Engineering Contradiction:
Improveuncertainty modeling capabilityVSAvoidenergy consumption
Core Design Contradiction:
ReliabilityVSUse of energy by moving object

Solution Approach 1:

The patent replaces complex probabilistic computation mechanisms with simpler analog hardware components. Specifically, it uses C-2C ladders for weight generation, jittery oscillators for randomness injection, and charge pumps for variance control, substituting digital probabilistic computations with analog circuit behaviors that naturally embody probability distributions.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent employs charge-based analog computing where electrical charge represents probability mass. The C-2C ladder structures use charge distribution to represent weight values, and charge pumps manipulate charge levels to control variance, leveraging electrostatic principles for efficient probabilistic computation.

Inventive Principle:
Principle #29Pneumatics and hydraulics

3Productivity

If analog computing-in-memory is used, then hardware efficiency and energy consumption are reduced, but precision and control accuracy may be compromised

Engineering Contradiction:
Improvecomputing efficiencyVSAvoidweight precision
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The patent segments the weight representation into discrete C-2C ladder units, where each unit contributes a quantized charge value. This segmentation enables precise analog weight representation through combinatorial charge summation, achieving high precision despite the analog implementation.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces jittery oscillators as intermediary components that add controlled randomness to the analog weight generation process. These oscillators modulate the charge output of C-2C ladders, enabling precise control over the variance of the probability distribution while maintaining the analog computing efficiency.

Inventive Principle:
Principle #24Intermediary (Mediator)

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 BNN implementation counteracts overfitting and sensitivity to attacks while being more efficient and cost-effective by introducing randomness and uncertainty, maintaining accuracy with smaller training data sets.

Implementation Method 1

a C-2C ladder to generate an output voltage corresponding to a mean weight value determined during training

Methodology Applied
Scientific EffectCapacitor charge transfer: Capacitance

Implementation Method 2

a jittery oscillator to generate a second clock signal

Methodology Applied
Scientific EffectJitter: Vibration

Implementation Method 3

a charge pump to adjust the output voltage based on a sample of the first clock signal

Methodology Applied
Scientific EffectCharge pumping: Pump

Data Source

PatentEP3989128B1An improved bayesian neural network and methods and apparatus to operate the same
Publication Date: 2025.08.13 INTEL CORP
  • EP3989128B1 patent drawingFigure 1
  • EP3989128B1 patent drawingFigure 2
  • EP3989128B1 patent drawingFigure 3

AI summary

Methods, apparatus, systems, and articles of manufacture providing an improved Bayesian neural network and methods and apparatus to operate the same are disclosed. An example apparatus includes an oscillator to generate a first clock signal; a resistive element to adjust a slope of a rising edge of a second clock signal; a voltage sampler to generate a sample based on at least one of (a) a first voltage of the first clock signal when a second voltage of the second clock signal satisfies a threshold or (b) a third voltage of the second clock signal when a fourth voltage of the first clock signal satisfies the threshold; and a charge pump to adjust a weight based on the sample, the weight to adjust data in a model.