Speaker
Description
We show that a quantum communication layer enables provable advantages throughout the
distributed learning pipeline, regardless of the model's nature (classical or quantum).
Starting from ring all-reduce (the universal primitive for large-scale gradient
aggregation) we prove that pre-shared entanglement reduces per-link \emph{online} communication by a
provably optimal factor of two and enables information-theoretic private aggregation
impossible for any classical protocol. Beyond bandwidth and security, the same quantum setup gives advantages directly
relevant to training: for margin-based
alignment testing (\textsc{GapIP}$_\tau$), the quantum advantage is \emph{quadratic}
($\widetilde{O}(\tau^{-1}\log P)$ qubits vs.\ $\widetilde{O}(\tau^{-2})$ bits); and for
sign-consistency auditing against private parameter matchings
(\textsc{TieAudit}$_\varepsilon$), it is \emph{exponential} ($O(\varepsilon^{-2}\log P)$
qubits vs.\ $\Omega(\sqrt{P})$ bits).
| Topic | Other |
|---|