Hierarchical Data Loading for ML Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Large-scale machine learning training faces inefficiencies due to data loading overheads when data exceeds local memory capacity, necessitating improved data storage and loading methods to enhance performance.
Innovation Solution
A hierarchical data loading method that divides original data into baseline and difference data, storing baseline data in faster memory devices and difference data in slower ones, allowing for efficient reconstruction and feeding into machine learning models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If data is stored in local memory (RAM/SSD), then access speed is fast, but memory capacity is insufficient for big data
Solution Approach 1:
The patent segments data into two distinct parts: baseline data (common across multiple datasets) and difference data (unique variations). This segmentation allows baseline data to be stored in fast local memory while difference data resides in slower but larger capacity storage, resolving the contradiction between access speed and memory capacity.
Solution Approach 2:
The patent introduces baseline data as an intermediary representation that captures commonalities across multiple datasets. This intermediary allows the system to store compressed difference data in large-capacity storage while maintaining fast access to the essential baseline information in local memory.
2Reliability
If all original data is loaded into memory for training, then training accuracy is maintained, but data loading overhead increases
Solution Approach 1:
The patent performs preliminary action by pre-computing and storing baseline data that represents common patterns across datasets. During training, this pre-prepared baseline can be quickly loaded and combined with smaller difference data, reducing the time required for data loading while maintaining training accuracy through faithful reconstruction of original data.
3Productivity
If data is divided and stored in different memory devices, then data loading efficiency improves, but system complexity increases
Solution Approach 1:
The patent applies local quality by assigning different storage locations based on data characteristics: baseline data (accessed frequently) is placed in fast local memory, while difference data (accessed less frequently) is placed in slower but larger capacity storage. This localized optimization improves data loading efficiency without requiring complex global memory management.
Data Source
AI summary
Original data for machine learning training can be received. The original data can be divided into baseline data and difference data. The baseline data and the difference data can be stored in different memory devices of the memory hierarchy associated with a computer, wherein the baseline data is stored in a first memory device having faster access speed than a second memory device in which the difference data is stored. The baseline data and the difference data can be loaded from the different memory devices. The original data can be reconstructed from the baseline data and the difference data. The reconstructed original data can be fed to a machine learning model to train the machine learning model.