Deep Neural Network Weight Quantization via Disruption Limits
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep neural network quantization techniques often compromise inference performance and robustness to reduce inference cost, making them unsuitable for edge devices with limited resources.
Innovation Solution
A method that determines a disruption limit value for each layer's weights in a deep neural network, allowing for targeted precision reduction without impacting inference performance, by adjusting weights based on a calculated adjustment limit value to minimize arithmetic precision while maintaining acceptable inference precision.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If the arithmetic precision of weights is reduced to decrease inference cost, then computing resources and energy consumption are reduced, but inference performance and robustness are drastically reduced
Solution Approach 1:
The patent applies different quantization strategies to different weights based on their individual disruption limit values. Instead of uniform quantization across all weights, each weight is analyzed separately and assigned an appropriate precision level, allowing critical weights to maintain higher precision while non-critical weights use lower precision, thus resolving the contradiction between inference cost and performance
Solution Approach 2:
The patent changes the arithmetic precision parameter of weights from a uniform value to variable values based on calculated disruption limit values. By dynamically adjusting precision levels for different weights according to their sensitivity analysis, the system achieves optimal balance between inference cost reduction and performance maintenance
2Device complexity
If uniform quantization is applied to all weights, then device complexity is reduced, but inference precision is significantly impacted
Solution Approach 1:
The patent segments the weight set into individual weight units, each subjected to separate disruption limit value calculation and quantization analysis. This segmentation allows precise control over each weight's quantization level, preventing the precision loss that would result from uniform quantization while managing complexity through automated per-weight analysis
Data Source
AI summary
The invention relates to a method for quantizing a deep neural network including several layers, previously trained during a training phase determining for each layer a set of weights. The method includes a phase of quantizing the deep neural network including determining a disruption limit value of at least one weight of the weight set of the layer, beyond which the output of the deep neural network is erroneous, determining, for a target inference precision of the neural network, and from the disruption limit value, an adjustment limit value of at least one weight of the set of weights, and decreasing an arithmetic precision of at least one weight of the set of weights as a function of the adjustment limit value. The invention also relates to a computer program, a device implementing such a method, and a deep neural network obtained by such a method.