Researchers have posted Meta-LSQ, a few-shot learning method that replaces MAML’s per-task gradient-descent loop with one regularized least-squares solve, on arXiv1.
MAML adapts a model to each new task with several gradient steps, and its outer training process must differentiate through those steps. Meta-LSQ instead fits a task using a fixed random Fourier basis and meta-learns three scalars: bandwidth, ridge and derivative weight. The approach could reduce task-level computation, while limiting the learned model to a fixed basis.
The paper’s abstract reports that Meta-LSQ trained in 14 seconds on 10-shot sinusoid regression and ended with 5.5 times lower error than a first-order MAML baseline that took 36 times longer to train. A Gaussian-process control suggested that sharing one kernel across tasks explained most of the gain, the authors said. These are the authors’ reported results, not an independent evaluation. This account draws on the arXiv abstract; the full paper text was not available in the retrieved record.
In simulated beam tuning at DESY’s ARES facility, Meta-LSQ was slightly more accurate than an ALPaCA-style model when given model derivatives, and performed well ahead of gradient descent, according to the abstract. In a separate simulation using measurements alone across the facility’s full range of target beams, its best setting had a median error of 7.3 micrometers, compared with 14.4 for the published final beams of a reinforcement-learning agent. The abstract gives no unit for 14.4. The agent had a lower median error on its 300 trials involving small targets.
On real sensor-calibration data, Meta-LSQ was ahead of MAML on two of four targets when the method used a learned prior mean. An exact Gaussian process with the same change and an ALPaCA-style model remained more accurate on most targets, the authors said. The abstract also says the method breaks down on nonsmooth targets and that generalized cross-validation can select too small a ridge for noiseless targets outside the basis span.
The arXiv record identifies the work as a preprint and says it was accepted at the Northern Lights Deep Learning Conference 2027, with publication planned in PMLR volume 352.