Promise and pitfalls of random projection algorithms for optical processors

Area of Science:

• * Emerging field of optical computing.
• * Application of random projection in artificial intelligence (AI).
• * High-dimensional data analysis and machine learning.

Background:

• * Random projection utilizes Gaussian i.i.d. matrices for efficient distance preservation in high-dimensional spaces.
• * Key machine learning algorithms leverage random projection for improved computational efficiency.
• * Optical implementations promise time and energy savings in computation.

Purpose of the Study:

• * Analyze algorithms employing random projection.
• * Evaluate the performance enhancement offered by optical random projection.
• * Investigate the computational time contribution of random projection within algorithms.

Main Methods:

• * Comprehensive analysis of algorithms including randomized singular value decomposition (RSVD), NEWMA, reservoir computing, and direct feedback alignment (DFA).
• * Examination of optical implementations of random projection.
• * Performance and computational time assessment of these methods.

Main Results:

• * Random projection constitutes a minor fraction of total computational time in analyzed algorithms.
• * Optical random projection does not yield significant performance enhancements.
• * Existing optical random projection initiatives show limited practical benefits.

Conclusions:

• * The computational advantage of random projection in optical computing is questionable.
• * Further research is needed to determine the true relevance of optical random projection units.
• * Current limitations suggest a need for re-evaluation of optical processing unit applications.