Vehicle Perception Neural Network Fixed-Point Quantization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing vehicle perception systems using neural networks face inefficiencies in processing resources due to the use of floating-point values, which can lead to increased power consumption and reduced performance in real-time vehicle control applications.
Innovation Solution
Implementing fixed-point quantization in neural networks for vehicle perception tasks, where a neural processing unit processes images captured by vehicle cameras, utilizing batch normalization and re-quantization techniques to convert floating-point values into 8-bit fixed-point values, reducing processing demands and enhancing efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If floating-point values are used in neural networks for vehicle perception, then measurement precision is improved, but use of energy increases and productivity decreases
Solution Approach 1:
The patent applies parameter changes by transitioning from floating-point representation to fixed-point representation with specific bit allocations (e.g., 8-bit integers, 16-bit fractions). This changes the numerical representation parameters while maintaining sufficient precision for vehicle perception tasks, thereby reducing computational complexity and power consumption without significantly degrading measurement precision
2Measurement precision
If floating-point values are used in neural networks for vehicle perception, then measurement precision is improved, but productivity decreases
Solution Approach 1:
The patent changes numerical representation parameters from floating-point to fixed-point formats with predetermined bit allocations. This parameter change enables faster arithmetic operations in neural network processing, improving productivity while maintaining adequate precision through careful selection of integer and fraction bit distributions
3Use of energy by moving object
If fixed-point quantization is implemented, then use of energy is reduced, but measurement precision deteriorates
Solution Approach 1:
The patent applies local quality by allocating different bit precisions to different components of the fixed-point representation based on their specific needs. Integer portions use fewer bits where coarse granularity suffices, while fraction portions use more bits where fine precision is required, optimizing the balance between energy consumption and measurement precision locally across different data types and computational stages
4Productivity
If fixed-point quantization is implemented, then productivity is improved, but device complexity increases
Solution Approach 1:
The patent applies segmentation by dividing the fixed-point number representation into distinct integer and fraction portions with specific bit allocations. This segmentation allows independent optimization of each portion's precision and enables modular implementation in hardware, managing device complexity through structured organization while maintaining high processing speed
Data Source
AI summary
Examples of techniques for using fixed-point quantization in deep neural networks are disclosed. In one example implementation according to aspects of the present disclosure, a computer-implemented method includes capturing a plurality of images at a camera associated with a vehicle and storing image data associated with the plurality of images to a memory. The method further includes dispatching vehicle perception tasks to a plurality of processing elements of an accelerator in communication with the memory. The method further includes performing, by at least one of the plurality of processing elements, the vehicle perception tasks for the vehicle perception using a neural network, wherein performing the vehicle perception tasks comprises quantizing a fixed-point value based on an activation input and a synapse weight. The method further includes controlling the vehicle based at least in part on a result of performing the vehicle perception tasks.