Complex-valued non-unitary ops boost QNN expressivity without depth: regression MSE drops from 9.55e-2 to 3.69e-4
bravo_abad · x · 2026-09-25
A thread on a paper by Wen and coauthors: instead of going deeper, adding a small number of imaginary-valued, non-unitary operations to conventional quantum circuits dramatically enlarges the space of transformations a quantum neural network can explore — without adding trainable parameters. In a regression task, MSE drops from 9.55×10⁻² to 3.69×10⁻⁴ at the same parameter count; for classification, a shallow complex-valued QNN reaches 96% accuracy, beating deeper conventional QNNs. The takeaway: transformation space, not parameter count, drives QNN capacity — a useful route for shallow, hardware-constrained circuits.
More from Research
- Senescent cell clearing therapy enters human trials this year — Kyrannio · 2026-10-24
- Interpretability demo shows chat and base models share the same "character neurons" — Sauers_ · 2026-10-24
- Steering a single neuron makes Qwen claim to be an angelic archangel AI — Sauers_ · 2026-10-24
- POLARIS-9B: a <$500, 4-GPU recipe for long-story writing that rivals 3x-larger models — MohitIyyer · 2026-10-24
- Seven independent tests suggest Pangram actually works as an AI text detector — TuhinChakr · 2026-10-24
- Why KV sharing works: 60-layer model keeps 21% of KV cache with lower loss — yoavartzi · 2026-10-24