Machine Learning at the Physical Limits: Energy, Accuracy and Quantum Security

The AI revolution has renewed interest in the fundamental limitations of computation set by physics. Thermodynamics imposes limitations on the energy cost of irreversible operations, while quantum mechanics imposes limitations on information access. A wave of in-physics AI mitigates these constraints by relaxing fixed-precision requirements and exploiting reversibility where the logic permits. 

The AI revolution has renewed interest in the fundamental limitations of computation set by physics. Thermodynamics imposes limitations on the energy cost of irreversible operations, while quantum mechanics imposes limitations on information access. A wave of in-physics AI mitigates these constraints by relaxing fixed-precision requirements and exploiting reversibility where the logic permits.

In this talk, I will present our advances in in‑physics machine learning, both in the radio‑frequency domain [1–2] and in the optical domain [3–5], that surpass the thermodynamic limitations of digital AI accelerators. I will then turn to optical computing for secure multiparty deep learning [6], where privacy is guaranteed by quantum mechanics.

1. Vadlamani*, Sulimany*,…, Englund. “Machine Intelligence on Wireless Edge Networks”. arXiv:2506.12210 (2025).

2. Gao, Vadlamani, Sulimany,…, Chen. “Disaggregated Deep Learning via In-Physics Computing at Radio Frequency”. arXiv:2504.17752 (2025).

3. Sulimany, Siago, Bandyopadhyay, Hamerly, Englund. “Photodetection Free Optical Machine Learning via Rectification”. CLEO (2025) (to appear).

4. Bandyopadhyay, Sulimany,…, and Englund “A Three-Terminal Nanophotonic Integrator for Deep Neural Networks” OFC, pp. 1-3. (2025).

5. Bacvanski, Vadlamani, Sulimany, and Englund. “QAMNet: Fast and Efficient Optical QAM Neural Networks” arXiv:2409.12305 (2024).

6. Sulimany, Vadlamani, Hamerly, Iyengar, Englund “Quantum-Secure Multiparty Deep Learning”. arXiv:2408.05629 (2024).