Quantized Classifier Training with Adjustable Distance Values
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Classifiers face a trade-off between classification performance and power consumption/memory requirements, particularly when deployed on embedded hardware, where reducing energy consumption and memory while maintaining performance is challenging.
Innovation Solution
A method for training classifiers that quantizes weights to different quantization types, allowing for reduced bit operations and efficient energy use, where the classifier is trained using a pre-defined quantization type to maintain performance and reduce parameter adjustments, enabling faster training with more data and adaptable resolution selection based on battery state or task difficulty.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the classifier uses high precision weights for classification, then classification performance is improved, but power consumption and memory requirements increase
Solution Approach 1:
The patent applies parameter changes by quantizing weights to different bit precisions (e.g., 8-bit, 4-bit, 1-bit) and adjusting the distance value threshold accordingly. This allows the classifier to operate at different precision levels, reducing computational complexity and power consumption while maintaining adequate classification performance for embedded hardware deployments.
2Measurement precision
If the classifier uses high precision weights for classification, then classification performance is improved, but memory requirements increase
Solution Approach 1:
The patent reduces memory requirements by quantizing weights from high precision (e.g., 32-bit floating point) to lower precision representations (e.g., 8-bit, 4-bit, or 1-bit integers). This parameter change significantly reduces the memory footprint of the classifier while maintaining adequate classification performance, making it suitable for embedded hardware with limited memory resources.
3Adaptability or versatility
If the classifier parameters are adjusted during training to different quantization types, then adaptability is improved, but training time increases
Solution Approach 1:
The patent performs preliminary action by pre-defining quantization types and their corresponding distance values before training. The distance value is adjusted in advance to compensate for quantization effects, allowing the classifier to be trained once with the adjusted distance value and then deployed at different quantization levels without requiring retraining. This significantly reduces training time while maintaining adaptability across different hardware platforms.
4Measurement precision
If the distance value is adjusted for training with quantized weights, then classification accuracy is improved, but computational complexity increases
Solution Approach 1:
The patent adjusts the distance value parameter to compensate for quantization effects, improving classification accuracy. While this adds a parameter adjustment step, the adjustment is performed once during training and then fixed, allowing the classifier to maintain high accuracy when deployed at different quantization levels without increasing operational computational complexity.
Data Source
AI summary
A computer-implemented method for training a classifier is disclosed. The classifier is designed to determine an output (y) for an input data point (x). The output (y) characterizes a classification of the input data point (x). The classifier comprises a multiplicity of weights on the basis of which the output (y) is determined. At least one weight of the multiplicity of weights is quantized to a predefined first number of first values. Each two consecutive first values differ by a distance value. The distance value is also adjusted for training and the multiplicity of weights is not adjusted.