Gradient Boosting Learning Device SRAM Buffer Management
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current gradient boosting technologies face limitations in processing speed and accuracy, particularly when handling large amounts of sample data, due to the limitations of static random access memory (SRAM) in storing sample data, which restricts the handling capacity and requires extensive hyperparameter adjustment and feature selection processes.
Innovation Solution
A learning device configured with a data storage unit, learning unit, update unit, subsampling unit, first buffer unit, and second buffer unit to manage learning data and gradient information, utilizing SRAM efficiently by buffering data in advance and writing it in blocks, enabling faster processing and handling of large sample data through optimized data management.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If SRAM is used to store sample data to shorten processing time, then processing speed is improved, but the amount of data that can be handled is limited
Solution Approach 1:
The patent divides the sample data into multiple blocks and processes them in segments. The buffer memory stores only the current block of data being processed, while other blocks are stored in external memory. This segmentation allows the system to handle large amounts of data by processing them in manageable chunks, resolving the contradiction between limited buffer capacity and large data volume requirements.
Solution Approach 2:
The patent performs preliminary actions by pre-processing and organizing data into blocks before loading them into the buffer memory. The system prepares data structures and organizes sample data in advance, so that when data needs to be processed, it is already in the optimal format and location, reducing the need for extensive hyperparameter adjustment and feature selection during the main processing phase.
2Measurement precision
If gradient boosting learning is performed multiple hundreds of times for hyperparameter adjustment and feature selection, then model accuracy and generalization performance are improved, but calculation time increases significantly
Solution Approach 1:
The patent performs feature selection and data preprocessing as preliminary actions before the main gradient boosting learning process. By selecting important features and organizing data in advance, the system reduces the dimensionality and complexity of the learning task, allowing for faster convergence and fewer iterations needed to achieve the desired model accuracy, thus reducing total calculation time.
Solution Approach 2:
The patent dynamically adjusts learning parameters such as subsampling rate and block size during the learning process. By changing these parameters adaptively based on progress and performance metrics, the system can achieve better generalization performance with fewer learning iterations, balancing accuracy improvement with calculation time constraints.
Data Source
AI summary
A learning device includes a learning unit configured to learn a model using learning data stored in a data storage unit; a subsampling unit configured to determine whether to use the learning data corresponding to gradient information updated by an update unit based on the model, for learning of a next model after learning of one model based on a subsampling rate; first and second buffer unit configured to buffer learning data determined to be used for and data determined not to be used for learning of the next model and gradient information corresponding to the learning data, respectively, up to a volume determined in advance. The first buffer unit and the second buffer unit are configured to write the learning data and the gradient information into the data storage unit for each predetermined block when buffering the learning data and the gradient information up to the volume.