Differences
This shows you the differences between two versions of the page.
| Both sides previous revisionPrevious revisionNext revision | Previous revision | ||
| projects:quantum:seqnn [2026/05/21 07:12] – [Roadmap position] kymki | projects:quantum:seqnn [2026/05/21 08:45] (current) – [Introduction] kymki | ||
|---|---|---|---|
| Line 105: | Line 105: | ||
| - | ====== | + | ====== |
| - | // | + | // |
| ===== Introduction ===== | ===== Introduction ===== | ||
| - | Kalman | + | Kalman |
| SEQNN, is an attempt to treat the parameters of a variational quantum model as latent states to be estimated sequentially, | SEQNN, is an attempt to treat the parameters of a variational quantum model as latent states to be estimated sequentially, | ||
| Line 132: | Line 132: | ||
| * each node maintains both an estimate and an uncertainty. | * each node maintains both an estimate and an uncertainty. | ||
| - | The important difference is not only "which update rule is used." | + | The important difference is not only "which update rule is used." |
| - | ===== Roadmap position ===== | + | ===== Roadmap position |
| ^ Stage ^ Status ^ | ^ Stage ^ Status ^ | ||
| Line 230: | Line 230: | ||
| $$ | $$ | ||
| - | In a supervised classifier this can look like a covariance-aware optimizer. In the generative experiments, | + | In a supervised classifier this can look like a covariance-aware optimizer. In the generative experiments, |
| ===== Model ladder ===== | ===== Model ladder ===== | ||
| - | The project | + | The project moved through increasingly demanding observation models. The key pattern is: do not add parameters, entanglement, |
| ^ Stage/model ^ Observation structure ^ What it tests ^ Claim status ^ | ^ Stage/model ^ Observation structure ^ What it tests ^ Claim status ^ | ||
| Line 243: | Line 243: | ||
| | Disjoint entangled '' | | Disjoint entangled '' | ||
| | Overlapping entangled '' | | Overlapping entangled '' | ||
| - | |||
| - | ===== Diagram: model progression ===== | ||
| - | |||
| - | < | ||
| - | <div style=" | ||
| - | <svg xmlns=" | ||
| - | < | ||
| - | <marker id=" | ||
| - | <path d=" | ||
| - | </ | ||
| - | </ | ||
| - | <rect x=" | ||
| - | <text x=" | ||
| - | <text x=" | ||
| - | <text x=" | ||
| - | <circle cx=" | ||
| - | <circle cx=" | ||
| - | <circle cx=" | ||
| - | |||
| - | <rect x=" | ||
| - | <text x=" | ||
| - | <text x=" | ||
| - | <text x=" | ||
| - | <path d=" | ||
| - | |||
| - | <rect x=" | ||
| - | <text x=" | ||
| - | <text x=" | ||
| - | <text x=" | ||
| - | <text x=" | ||
| - | |||
| - | <rect x=" | ||
| - | <text x=" | ||
| - | <text x=" | ||
| - | <text x=" | ||
| - | <line x1=" | ||
| - | <line x1=" | ||
| - | |||
| - | <rect x=" | ||
| - | <text x=" | ||
| - | <text x=" | ||
| - | <text x=" | ||
| - | <circle cx=" | ||
| - | <circle cx=" | ||
| - | <circle cx=" | ||
| - | <circle cx=" | ||
| - | <line x1=" | ||
| - | <line x1=" | ||
| - | <line x1=" | ||
| - | <text x=" | ||
| - | <text x=" | ||
| - | <text x=" | ||
| - | <text x=" | ||
| - | |||
| - | <line x1=" | ||
| - | <line x1=" | ||
| - | <line x1=" | ||
| - | <path d=" | ||
| - | <text x=" | ||
| - | </ | ||
| - | </ | ||
| - | </ | ||
| ===== The Stage 5 observation model ===== | ===== The Stage 5 observation model ===== | ||
| Line 366: | Line 304: | ||
| ===== Models compared in Stage 5 ===== | ===== Models compared in Stage 5 ===== | ||
| - | ^ Model ^ Role ^ State/ | + | ^ Model ^ Role ^ State/ |
| | Adam | optimizer baseline | point estimate, no posterior covariance | whether a standard adaptive optimizer tracks the same sampled streams well | | | Adam | optimizer baseline | point estimate, no posterior covariance | whether a standard adaptive optimizer tracks the same sampled streams well | | ||
| | Gradient descent | optimizer baseline | point estimate, no posterior covariance | whether a tuned first-order method is enough | | | Gradient descent | optimizer baseline | point estimate, no posterior covariance | whether a tuned first-order method is enough | | ||
| Line 467: | Line 405: | ||
| ===== What remains for the Stage 5 publication path ===== | ===== What remains for the Stage 5 publication path ===== | ||
| - | The remaining work is about strengthening the evidence | + | The remaining work is about strengthening the evidence, not changing the core idea. |
| - Decide whether to run a 50-seed sensitivity confirmation. | - Decide whether to run a 50-seed sensitivity confirmation. | ||
| Line 482: | Line 420: | ||
| * It helps show where distributed Fisher factors remain stable or attractive as the chain grows. | * It helps show where distributed Fisher factors remain stable or attractive as the chain grows. | ||
| - | ===== What we are not claiming | + | ===== What I am not claiming ===== |
| - | This is important for the eventual paper framing. | ||
| - | |||
| - | * Not a quantum advantage claim. | ||
| * Not a hardware claim; these are controlled generative simulations. | * Not a hardware claim; these are controlled generative simulations. | ||
| * Not an arbitrary-entangled-circuit claim. | * Not an arbitrary-entangled-circuit claim. | ||
| Line 494: | Line 429: | ||
| * Not a claim that every SEQNN variant works; the independent EKF is intentionally a negative control in Stage 5. | * Not a claim that every SEQNN variant works; the independent EKF is intentionally a negative control in Stage 5. | ||
| - | The narrower claim is stronger: | + | Rather: |
| > In an identifiable overlapping entangled generative circuit, a distributed SEQNN estimator using Fisher/ | > In an identifiable overlapping entangled generative circuit, a distributed SEQNN estimator using Fisher/ | ||
| - | ===== Publication framing ===== | ||
| - | |||
| - | A good paper story is emerging: | ||
| - | |||
| - | - Start with the problem: VQC training usually treats circuit parameters as optimizer variables, not latent states. | ||
| - | - Introduce SEQNN: local EKF estimators attached to quantum circuit parameters. | ||
| - | - Distinguish supervised training from true generative estimation. | ||
| - | - Show the model ladder from local identifiable circuits to richer local observations. | ||
| - | - Use Stage 5 as the main contribution: | ||
| - | - Compare against Adam, GD, SPSA, independent SEQNN-EKF, and centralized EKF. | ||
| - | - Be explicit that Fisher factors are the promoted distributed architecture. | ||
| - | - Leave topology diffusion and augmented noise-state EKF as future work. | ||
| - | |||
| - | The paper claim should be written as exploratory but concrete. The work is not finished because the scaling slice is still needed and the 50-seed sensitivity decision remains open. But Stage 5 now has a defensible center: the overlapping '' | ||
| - | |||
| - | ===== Suggested abstract-style summary ===== | ||
| - | |||
| - | Self-Estimating Quantum Neural Networks frame variational quantum circuit training as sequential Bayesian estimation rather than global optimization. In the current generative benchmark line, circuit parameters are latent states, finite-shot measurements are observations, | ||