On quantum backpropagation, information reuse, and cheating measurement collapse

Item: On quantum backpropagation, information reuse, and cheating measurement collapse
Rating: 5
Author: PostSingularity Institute

by QAISG · View the channel on PSI

Formal & Physical Sciences AI — Models & Research UYQArtificial intelligence (AI)UYQMMachine learning

🎙 QAISG 👥 397 📅 September 6, 2023 ⏱ 48 min 👁 137 📄 original study 🧭 2026-10-09

Available in: English (current) Français

Keywords

quantum backpropagationinformation reusemeasurement collapseshadow tomographygradient scaling

Summary

The talk presents a theoretical analysis of backpropagation in quantum machine learning models. Given a parameterized quantum circuit, the goal is to compute gradients efficiently for optimization. The speaker contrasts classical backpropagation, which reuses intermediate information to achieve runtime proportional to a single forward pass, with the naive quantum approach that incurs a quadratic overhead in the number of parameters due to measurement collapse. The talk first shows that without quantum memory (i.e., single-copy access), achieving backpropagation scaling is impossible even with known input states, and existing methods like the parameter-shift rule have O(m^2) cost. It then demonstrates that with access to multiple copies of the state (quantum memory), one can design algorithms based on shadow tomography that achieve nearly linear scaling for restricted classes of operators (e.g., Pauli operators). More generally, the problem of estimating all gradients is shown to be connected to shadow tomography, highlighting the inherent difficulty of reusing quantum information. The work clarifies the nuances of information reuse in quantum settings and the fundamental challenges in training large quantum models.

174 words

Critical Evaluation

Value of the Information & Strength of the Argument

The value of the information is high, as it addresses a critical bottleneck in quantum machine learning—the computational cost of gradient estimation—and provides both impossibility results and constructive algorithms. The argumentation is rigorous and well-structured: the speaker first defines the cost model and clearly explains the quadratic scaling of existing methods, then systematically explores different settings (known vs. unknown input states, single-copy vs. multi-copy access). The connection to shadow tomography is insightful and places the problem in a broader theoretical context. The presentation includes intuitive examples and actively engages the audience to check understanding, reinforcing the clarity of the reasoning.

Scientific Rigor, Source Quality, Title Accuracy

The scientific rigor is excellent: the talk is based on a preprint (arXiv:2305.13362) and references prior work on shadow tomography and gradient methods. The claims are carefully derived, with explicit mention of assumptions and lower bounds. The title accurately represents the content, and the presentation adheres closely to the paper’s results. No external sources beyond the paper and a few cited works are mentioned, but the description provides the arXiv link for verification. The audience’s reactions (only 2 likes) do not provide enough data for comment analysis, so no trends can be inferred.

208 words

Title / Content Match

The title perfectly reflects the content: the talk addresses quantum backpropagation, information reuse, and the role of measurement collapse in quantum gradient computation.

Quality & Reliability

9/10

The content is a rigorous research presentation of a peer-reviewed paper (arXiv:2305.13362), with clear technical arguments and references to related work. The speaker demonstrates deep expertise and the theoretical results are well-motivated.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

On quantum backpropagation, information reuse, and cheating measurement collapse — The paper presented in the talk, containing the full results and proofs.

Concurring Sources

On quantum backpropagation, information reuse, and cheating measurement collapse — The primary source, consistent with the talk's content.

Contribution & Novelties

The talk presents original research that clarifies the resource requirements for gradient computation in quantum machine learning. It shows that without quantum memory, backpropagation scaling is impossible, and with multi-copy access, a shadow-tomography-based algorithm achieves near-optimal scaling for Pauli operators. The connection to shadow tomography provides a deeper understanding of information reuse in quantum systems.

Pour aller plus loin :

Quantum machine learning — Overview of quantum ML models and their training.
Backpropagation — Classical gradient computation algorithm and its efficiency.
Shadow tomography — Aaronson’s original paper on shadow tomography of quantum states.
Parameter shift rule — Common method for computing quantum gradients, with quadratic scaling.

106 words

Radar Profile

The radar profile shows extremely high scores across all dimensions, with a strong peak in technical level (10/10), reflecting the advanced quantum computing concepts. The information quality and quantity are also near-perfect, consistent with a focused research presentation.

Reliability 10/10