arXiv:2610.34055
Quantum neuroplastic learning for time-series processing
Mostafizur Rahaman Laskar
Abstract
Quantum machine learning uses quantum states and dynamics to represent and process data. In quantum reservoir computing, the physical system is typically fixed and a classical readout is trained. Inspired by the formation, strengthening and pruning of synapses during biological learning, we propose quantum neuroplastic learning (QNL), which reorganizes the couplings among the qubits of a quantum reservoir, much as experience rewires connections among neurons. Eligibility traces guide local edits to couplings, fields, gate angles and evolution time, and an edit is retained when a statistical bound on its improvement exceeds a complexity cost. We evaluate QNL under restarting and online monitored protocols on short- and long-term memory, nonlinear autoregressive processing and forecasting, and channel equalization tasks. In simulation, QNL enhances memory capacity, develops long-term recall and sustains adaptation across multiple tasks, with the largest benefit for tasks that require extended memory. On the IBM superconducting processor \texttt{ibm\_fez}, with six seven-qubit reservoirs running in parallel on 42 qubits (1.28 million shots in total), configurations learned in simulation and frozen before execution reduced the channel-equalization symbol error rate from 9.7\% to 2.2\% and achieved a lower error rate than a matched-budget tuned baseline.
AI Summary
Can editing a quantum reservoir’s physical configuration improve temporal processing beyond training its readout? Quantum neuroplastic learning (QNL) combines a refitted linear readout with gradient-free edits to couplers, fields, gate angles and evolution time. Decaying eligibility traces bias proposals; an edit is accepted when a nominal Student-$t$ lower confidence bound on paired loss reduction exceeds a cost for added couplers. Qubit count, measurement protocol and readout architecture remain fixed. Simulations of five-qubit monitored reservoirs show summed short-term-memory capacity gains of $4.92$–$5.94$ with exact expectation values. Training for delays $15$–$30$ raises long-term-memory capacity from near zero to $3.46$ under restarting and $3.40$ under weak monitoring, but only $0.62$ under stronger monitoring. Adaptation continues across 16 task switches; NARMA forecasting errors decrease descriptively by approximately $10$–$20\%$. Channel-equalization simulations reach errors comparable to random search over the same candidate count, with QNL requiring $25\%$ more stream evaluations. Seven-qubit configurations learned in simulation and frozen before IBM hardware execution reduce symbol error from $9.7\%$ to $2.2\%$ in an initial run. A chain-balanced comparison gives $4.2\%$, versus $5.8\%$ for tuned random search and $13.0\%$ for the reference, using 600 rather than 720 CZ gates per message. Limitations include small reservoirs, few realizations, two runs on one device, and no test of earlier-task retention. Hardware learning required approximately $10^9$ simulated shots. Exact-feature memory gains required large measurement ensembles; noise models omit effects including gate errors and crosstalk, and admission bounds rely on approximate normality.
Error found: The abstract's 'matched-budget tuned baseline' matches candidate counts but excludes QNL's incumbent reevaluations, which give QNL 25% more learning-stream evaluations, so the hardware results do not establish superiority at equal total learning budget.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- This is a first draft of the work on ideation. The data, results, figures and equations may be revised based on our ongoing work, and revision from Journal. Upon acceptance in a Journal, the arxiv version may also be withdrawn based on Journal's policy