Wits researchers build optical neural network that reads light without imaging

Researchers at Wits University have built an optical neural network that identifies the topology of structured light patterns (optical skyrmions) without capturing a detailed image. The system recognised 81 structures spanning 15 skyrmion numbers with about 93% accuracy and held 100% accuracy under noise up to 95% of the signal. The work was published in Nature Communications on 7 October.

• Network identifies 81 light structures across 15 skyrmion numbers at ~93% accuracy
• 100% accuracy maintained with noise up to 95% of signal intensity
• Training parameters cut by a factor of about 1,000
• Only about 15% of light reaches the detector due to losses