🔗 Source: arXiv

Streamlined optical training of large-scale modern deep learning architectures with direct feedback alignment

🚀 Technical Novelty

  • Mechanism: Implements Direct Feedback Alignment (DFA) on a hybrid electronic-photonic platform where an Optical Processing Unit (OPU) passively executes large-scale random matrix multiplications in parallel at up to 1500 TeraOPS under 30W.
  • Nuance: Replaces the sequential, backward-locking nature of backpropagation with fully parallel layer updates via optical random projections, decoupling training time from electronic compute/memory bottlenecks and enabling superior scaling for ultra-deep/wide networks.

💡 Yield

  • Successfully trained >1B parameter Transformers and Diffusion models across language, vision, and generative tasks; demonstrated ODFA surpasses backpropagation in wall-clock training time at extreme scales (~2.7B parameters) while maintaining high energy efficiency.

⚠️ Limitations

  • Current optical projection dimensions (10³ × 10³) significantly underutilize the hardware’s maximum capacity (10⁶ × 10⁶); GPU memory constraints still limit forward passes and parameter updates for massive models, requiring offloading techniques; smaller networks initially train slower due to fixed optical frame rate limits.