Random Projection Base Vector Generation Using Seeds
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional random projection techniques face challenges in speed and efficiency due to high data transfer costs and access delays associated with transferring base vectors across a system bus, especially in low-end systems with limited bandwidth.
Innovation Solution
An information processing apparatus and method that generates base vectors using random number seeds stored in memory, reducing the need to transfer entire base vectors by dynamically generating them within the system, thereby reducing data transfer requirements and improving processing speed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If base vectors are stored in external memory and transferred via system bus, then projection operation can be performed, but data transfer time exceeds processing time due to limited bandwidth and access delay
Solution Approach 1:
The base vectors are segmented into two parts: random number seeds (small size) stored in external memory, and actual base vectors (large size) generated dynamically in accelerator hardware. This segmentation allows transferring only the small seed data while generating the large base vectors locally, dramatically reducing transfer time and bandwidth requirements.
Solution Approach 2:
The random number seeds are prepared and stored in external memory in advance before the projection operation. These seeds serve as compact templates that enable rapid generation of full base vectors in hardware, eliminating the need to transfer complete base vectors and thus reducing data transfer time.
2Adaptability or versatility
If random number generator is incorporated in accelerator hardware, then base vectors can be generated dynamically, but this approach cannot be applied when selected random number sequences are required
Solution Approach 1:
Random number seeds serve as intermediaries between external memory and the random number generator in accelerator hardware. These compact seeds enable the hardware generator to produce specific, selectable random number sequences without requiring the system to store or transfer complete base vectors, thus maintaining adaptability while simplifying the system.
3Productivity
If all base vectors are formed from random number sequences, then implementation method with random number generator can be adopted, but selected random number sequences cannot be utilized
Solution Approach 1:
The system changes the parameter representation from complete base vectors to compact random number seeds. This parameter transformation enables both high processing efficiency (through hardware-based dynamic generation) and adaptability (through selective seed storage and controlled generation of specific random sequences).
Data Source
AI summary
A plurality of random number sequences are generated using a plurality of random number seeds stored in a memory, and a plurality of base vectors are generated based on the plurality of random number sequences. A transformation source vector is transformed into a lower-dimensional vector by performing random projection for the transformation source vector using the plurality of generated base vectors.