Short Industry Talks pt. 1
summary
Implements a method to estimate the canonical partition function by interpolating outcomes from several Trotter step sizes to extrapolate zero-step exponential operators, avoiding block encoding and saving about 20–30% qubits for a Majorana-rich model; uses Chebyshev interpolation and amplitude estimation to recover Boltzmann traces with polylogarithmic systematic-error scaling. Proposes a training-time privacy test called plausible deniability that adds noise to batch gradients, rejects gradients without T similar noisy neighbors, and yields higher utility than DP-SGD while admitting weaker formal DP parameters.
AI-written summary · about
chapters (4)
related talks
01h 14mTutorial: Privacy, Part III
43mPrivacy Amplification by Sampling in Practice
46mLearning the Privacy-Utility Trade-off with Bayesian Optimization
01h 14mTutorial: Privacy, Part II
37mUser-level and federated local differential privacy
31mScaling privacy embedded distributed ML systems
54mPrivacy Amplification from Structured Algorithmic Randomness
47mPrivate targeting strategies