Deep Learning Machine Adaptive Integer Quantization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning algorithms based on integer operations suffer recognition performance deterioration when processing unprocessed data, especially when data is concentrated at extreme ends, leading to decreased performance in systems with limited resources like smartphones.
Innovation Solution
A deep learning machine and operation method that classify input data into two types, integerize data requiring conversion, and infer output using a classification network combined with the deep learning model, maintaining performance and accuracy by using labeling criteria derived from inference performance values and quantized learned data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If integer operations are used in deep learning to reduce computational complexity, then the number of operations is reduced and processing speed is improved, but recognition performance deteriorates due to feature loss during quantization
Solution Approach 1:
The patent applies parameter changes by dynamically adjusting the quantization parameters based on the classification results. When input data is classified as requiring high precision, the system uses higher precision integer operations (e.g., 32-bit integers) instead of lower precision (e.g., 8-bit integers), thereby adapting the computational precision to the actual needs of each input sample and minimizing recognition performance deterioration.
Solution Approach 2:
The patent implements dynamics by making the integer operation precision adaptive rather than static. The system dynamically switches between different precision levels based on the classification output, allowing the deep learning model to operate at high precision when necessary and at lower precision when acceptable, thus resolving the contradiction between processing speed and recognition performance.
2Measurement precision
If floating point operations are used to maintain recognition performance, then accuracy is preserved, but the number of operations increases and processing speed decreases
Solution Approach 1:
The patent applies segmentation by dividing the input data processing into two segments: first, a classification stage that quickly categorizes input data, and second, an inference stage that applies integer operations with adaptive precision. This segmentation allows the system to avoid full-precision floating point operations for all inputs, thereby improving processing speed while maintaining recognition performance for critical cases.
Solution Approach 2:
The patent uses partial action by applying high-precision operations only to the extent necessary - specifically, only to input data that the classification stage identifies as requiring high precision. For other inputs, lower precision integer operations suffice, reducing the overall computational burden while maintaining adequate recognition performance.
3Ease of manufacture
If data is concentrated at extreme ends during quantization, then integer operations become more efficient, but feature loss increases and recognition performance deteriorates
Solution Approach 1:
The patent applies preliminary action by performing classification on the input data before the quantization and integer operation stages. This preliminary classification identifies which inputs will benefit from integer operations and which require higher precision, allowing the system to prepare appropriate processing paths in advance and avoid feature loss for critical inputs.
Solution Approach 2:
The patent introduces an intermediary classification model that acts as a mediator between the input data and the integer operation-based deep learning model. This intermediary assesses the input data and determines the appropriate precision level, preventing feature loss by routing suitable inputs to integer operations while protecting sensitive inputs from aggressive quantization.
Data Source
AI summary
A deep learning machine includes a classification unit having a labeling criterion and configured to label input data according to the labeling criterion, a conversion unit configured to integerize input data labeled as a first type requiring integerization among the input data labeled by the classification unit, a first learning data unit configured to receive the input data of the first type integerized through the conversion unit and to infer output data, and a second learning data unit configured to receive input data labeled as a second type requiring no integerization and to infer the output data.