Tag: quantum-machine-learning
All the articles with the tag "quantum-machine-learning".
-
On Quantum Algorithms for Neural-Network Training: A Conjecture on Fault-Tolerant Speedups
The Suhubdy Conjecture — a three-form mathematical conjecture for a fault-tolerant quantum algorithm that accelerates — or replaces — backpropagation for empirical-risk minimisation in neural networks. Positioned against the Baur–Strassen cheap-gradient bound, Jordan and Gilyén–Arunachalam–Wiebe gradient-estimation algorithms, Abbas et al. (NeurIPS 2023) impossibility, quantum SDP and interior-point solvers, the Zhang–Leng–Li and Garg–Kothari–Netrapalli–Sherif lower bounds, and the 2024–2026 line on quantum non-logconcave sampling. Companion arXiv-style preprint (.tex source and .pdf) provided.
-
The Credit Assignment Problem: From Rosenblatt's Perceptron to Backpropagation to Quantum Gradients
A long-form, citation-grounded history of the credit assignment problem — the core question of how to apportion blame for a global error across the internal parameters of a learning machine. From Rosenblatt's 1958 perceptron rule, through Linnainmaa, Werbos, and the 1986 Rumelhart-Hinton-Williams paper, to modern alternatives (feedback alignment, equilibrium propagation, predictive coding, forward-forward, synthetic gradients), reinforcement-learning credit assignment, and the frontier of quantum gradients (parameter-shift rule, quantum natural gradient, HHL, and the Abbas et al. NeurIPS 2023 impossibility result).
-
On Quantum Backpropagation, Information Reuse, and Cheating Measurement Collapse
Can quantum models train as efficiently as classical neural networks? Abbas et al. prove backpropagation scaling is impossible without multiple copies of a quantum state, then show how shadow tomography gets you most of the way there. A deep dive into the paper that defines the trainability frontier for quantum machine learning.