arXiv:2610.11360
When Does Interference Help Learning? Kernel Geometry as a Pre-Experimental Test for Photonic Reservoir Computing
Oishik Kar, Aswath Babu H
OISHIK KAR IIIT DharwadAbstract
Multi-photon interference is the resource most native to photonic quantum machine learning, yet experiments on photonic reservoirs have reported advantages, null results, and train-only effects. These outcomes cannot be compared: each measures task-dependent accuracy on different tasks and platforms. Here we resolve the question task-independently through the kernel geometry of boson-sampling feature maps. Interference redistributes feature variance across roughly twice as many effective dimensions as distinguishable-particle statistics, generating a large geometric separation between the corresponding kernels. The separation grows superlinearly with indistinguishability, is linear in Hong--Ou--Mandel visibility at small visibility for two to four photons, and retains approximately 80% of its magnitude at the source quality of current quantum processors. Under experimental sampling budgets the separation is learnable on adversarially constructed tasks (accuracy advantage +0.22 plus or minus 0.05) and an order of magnitude smaller on natural tasks, gated by kernel--task alignment. The results reconcile the existing experimental record and supply a pre-experimental protocol: kernel geometry establishes that an interference-specific learning resource exists, and task alignment determines whether a given task can access it; both are computable before committing significant hardware resources.
AI Summary
The paper asks when multiphoton interference improves learning in photonic reservoirs compared with distinguishable-photon statistics through the same interferometer. It studies kernels formed from collision-free output probabilities, separating task-independent kernel geometry from task-dependent kernel–target alignment. Exact classical simulations use Haar-random interferometers, phase-encoded inputs, two to four photons, and partially distinguishable photon statistics. At the main geometry configuration—three photons, 12 modes, and 100 inputs—the indistinguishable kernel reaches 90% cumulative spectral weight in approximately 61 dimensions versus 33 for the distinguishable kernel. The geometric difference is $g=10.3\pm1.1$ across ten interferometers. Its excess above one grows linearly with Hong–Ou–Mandel visibility near zero; at visibility $0.864$, approximately 80% of the full-visibility excess remains. Kernel ridge classification on deliberately constructed, out-of-sample-labeled tasks gives an accuracy advantage of $0.222\pm0.046$. Natural-task advantages are smaller: approximately $0.024$–$0.041$ for several synthetic families and $0.036\pm0.054$ for digit classification. Reconstructing kernels and designed labels under multinomial sampling preserves the advantage at $3\times10^4$ shots per input. These observations motivate screening candidate experiments using geometry and task alignment. The stated limitations include small, classically simulable systems, synthetic adversarial tasks, conditioning-sensitive geometric differences, retrospective experimental comparisons, and no modeling of temporal feedback or multiphoton emission noise. A preregistered expanded task battery did not validate fine-grained alignment-based predictions: rank correlation was $0.027$ ($p=0.93$), with sign agreement in 9 of 13 families. Hardware validation remains outstanding, and classical kernel computation becomes prohibitive at large photon numbers.
Central claim may not hold: The abstract presents kernel–task alignment as determining task accessibility, although the independent natural-task battery fails both registered predictive tests, leaving the proposed task-selection rule unvalidated. The shot-noise protocol uses exact probabilities for outcome trimming despite claiming sampled-data-only reconstruction, and its changing labels do not test robustness of a fixed learning task. The specified probability normalization does not give the two raw kernels identical traces, so the claimed equal-total-variance spectral comparison requires an undocumented normalization or correction. Collision-free renormalization generally destroys the affine two-photon feature mixture assumed in Supplementary Note 7, and its quadratic-kernel expansion also contains an incorrect second-order coefficient. The paper substitutes y^T K y for inverse-kernel learning complexity and uses its convexity to exclude geometric interior optima, a conclusion that does not follow.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 15 pages 5 figures