Parameterized Weight Clipping for Binarized Neural Network Accuracy

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Binarized neural networks (BNNs) face challenges in accuracy due to aggressive parameter reduction, leading to significant computational and memory efficiency gains but with considerable accuracy loss, and existing methods to improve performance through parameter processing or modulation are limited.

Innovation Solution

The method involves training BNNs using a parameterized weight clipping (PWC) scheme, where a clipping function is applied to generate a binarized weight set, and the range of the clipping function is adaptively updated based on gradients during the backpropagation process to minimize accuracy loss and address gradient mismatch issues.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If parameters are aggressively reduced to 1-bit in binarized neural networks, then memory footprint and computational speed are significantly improved, but accuracy is considerably lost

Engineering Contradiction:
Improvecomputational speedVSAvoidaccuracy
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The patent applies parameterized weight clipping where the clipping range is dynamically adjusted based on the magnitude of weights during training. This allows the system to adaptively change the parameter range of the clipping function, enabling better preservation of important weight information while maintaining the binarized structure, thus improving accuracy without sacrificing the computational efficiency of 1-bit weights

Inventive Principle:
Principle #35Parameter changes

2Quantity of substance

If parameters are aggressively reduced to 1-bit in binarized neural networks, then memory footprint is significantly reduced, but accuracy is considerably lost

Engineering Contradiction:
Improvememory footprintVSAvoidaccuracy
Core Design Contradiction:
Quantity of substanceVSMeasurement precision

Solution Approach 1:

The patent dynamically adjusts the clipping range parameters during training based on weight magnitude statistics. This adaptive parameter change allows the system to optimize the balance between compression ratio and accuracy preservation, maintaining significant memory footprint reduction while improving accuracy through intelligent parameter modulation

Inventive Principle:
Principle #35Parameter changes

3Measurement precision

If parameter processing or modulation is applied to improve BNN performance, then accuracy may be improved, but the complexity of training increases

Engineering Contradiction:
ImproveaccuracyVSAvoidtraining complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent introduces parameterized weight clipping with dynamically adjustable range parameters that are updated during training based on gradient computation. While this adds some training complexity compared to fixed clipping, the parameter updates follow a straightforward gradient-based optimization process that integrates naturally with existing training frameworks, providing a manageable increase in complexity for significant accuracy improvements

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent implements a feedback mechanism where the clipping range parameters are continuously adjusted during training based on the computed gradients and weight magnitude statistics. This feedback loop enables the system to automatically adapt the clipping parameters to the current state of the network, improving accuracy through iterative optimization without requiring manual intervention or complex external control systems

Inventive Principle:
Principle #23Feedback

Data Source

PatentUS20230351189A1Method of training binarized neural network with parameterized weight clipping and memory device using the same
Publication Date: 2023.11.02 SAMSUNG ELECTRONICS CO LTD
  • US20230351189A1 patent drawing
  • US20230351189A1 patent drawing
  • US20230351189A1 patent drawing

AI summary

In a method of training a binarized neural network (BNN), a binarized weight set is generated by applying a clipping function to a weight set. Output data is generated by sequentially performing a forward computation on the binarized neural network based on input data and the binarized weight set. A gradient of the weight set is generated by sequentially performing a backward computation on the binarized neural network based on loss calculated from the output data. The binarized neural network is trained by updating the weight set based on the gradient of the weight set and changing a range of the clipping function.