Machine Intelligence at the Wireless Edge

Machine intelligence on wireless edge networks promises low-latency, low-power inference, yet deployment remains constrained by memory footprints, energy for data movement and conversion, limited bandwidth, and security risks. We introduce MIWEN [1]: instead of offloading queries, a base station broadcasts model weights while clients perform inference through in-physics computation within the receive chain. This disaggregated model access via wireless broadcasting decouples model capacity from client memory and eliminates repeated signal conversions. We analyze the resulting energy–accuracy trade-offs and identify an optimal operational window. Companion experiments [2] demonstrate complex-valued matrix–vector multiplication, achieving 95.7 % accuracy at 6.0 fJ/MAC (165.8 TOPS/W) per client. While our protocol inherently preserves client-data privacy, we show that double-blind operation is further achievable [3]. Together, these results pave the way to scalable, memory-free, energy-efficient, and secure edge AI.

Supported by the European Union’s Horizon programme under Marie Skłodowska-Curie Grant No. 101202109.

[1] Krishna*, Sulimany* et al., arXiv:2506.12210 (2025) [2] Zhihui et al., arXiv:2504.17752 (2025) [3] Sulimany et al., arXiv:2408.05629 (2024)