Electronic Apparatus with Dataset-Aware AI Preloader Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Cloud platforms experience bottlenecks and inefficiencies in AI model training due to network delays, dependency issues between processors and memory, and wasted storage space from unconsidered locality of cache usage, leading to reduced job efficiency and expandability.
Innovation Solution
An electronic apparatus and method that utilizes a processor to identify the type of preloader based on dataset characteristics, transmitting either a first type that copies data or a second type that generates links, optimizing data distribution to computing devices for efficient AI model training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If learning data is stored in local memory inside a computing device, then training speed is improved, but dependency between processor and memory occurs and expandability is degraded
Solution Approach 1:
The patent introduces a data manager as an intermediary component that mediates between the processor and storage devices. The data manager handles data loading, caching, and coordination across multiple storage devices, eliminating direct processor-memory dependencies while maintaining fast access speeds. This intermediary layer enables flexible data management without hardcoding specific storage configurations.
Solution Approach 2:
The patent implements dynamic data loading strategies where the system can adaptively select which storage devices to access based on real-time needs. The data manager dynamically manages data cached in each storage device, allowing the system to flexibly adjust data access patterns without fixed processor-memory bindings, thereby improving expandability while maintaining performance.
2Adaptability or versatility
If network transmission is used to load learning data, then data accessibility is improved, but network delay causes bottleneck and training speed decreases
Solution Approach 1:
The patent implements a preloader mechanism that performs preliminary data loading actions before the actual training begins. The preloader loads learning data into local storage devices or cache memory in advance, so when training starts, the data is already available locally without network delay. This preliminary action eliminates the network bottleneck during the critical training phase while maintaining data accessibility through the same network infrastructure.
3Speed
If local cache is provided in each computing device, then data access speed is improved, but locality is not considered causing overlapping and storage space waste
Solution Approach 1:
The patent introduces a feedback mechanism where the data manager monitors which storage devices are accessed most frequently during training. Based on this feedback, the system dynamically adjusts data distribution strategies, loading data into storage devices with higher access patterns. This feedback loop prevents redundant data caching across multiple devices while optimizing storage space utilization based on actual access needs.
Solution Approach 2:
The patent implements local quality optimization by considering the specific access patterns and locality characteristics of different datasets when distributing data to storage devices. Instead of uniform caching, the system tailors data distribution to match actual training needs, loading data into appropriate storage devices based on their accessibility and capacity characteristics, thereby reducing redundant storage while maintaining fast local access.
4Productivity
If multiple computing devices train using the same learning data, then training efficiency is improved, but dependency on data availability increases
Solution Approach 1:
The preloader mechanism performs preliminary data loading actions for multiple computing devices before they begin training. By loading data into local storage or cache in advance, the system ensures data availability is established beforehand, reducing dependency on real-time data availability during parallel training operations. This enables multiple devices to work independently with confidence that their data is prepared.
Data Source
AI summary
An electronic device may include a communicator and a processor configured to, based on characteristic information of a dataset for performing a job included in a queue, identify a type of a preloader configured to preload the dataset, and control the communicator to transmit the preloader of the identified type to at least one computing device among a plurality of computing devices. The job may be related to performing training of an AI model by using the dataset. The preloader of a first type may copy the dataset stored in at least one external storage device and store the dataset in the at least one computing device, and the preloader of a second type may generate a connecting link for the dataset stored in the at least one external storage and store the connecting link in the at least one computing device.