Search

In practice, we usually need to build variable-sized models adapting for diverse resource constraints in different application scenarios, where weight initialization is an important step prior to training. The Learngene framework, introduced recently, firstly learns one compact part termed as learngene from a large well-trained model, after which learngene is expanded to initialize variable-sized models. In this paper, we start from analysing the importance of guidance for the expansion of well-trained learngene layers, inspiring the design of a simple but highly effective Learngene approach termed SWS (Stage-wise Weight Sharing), where both learngene layers and their learning process critically contribute to providing knowledge and guidance for initializing models at varying scales. Specifically, to learn learngene layers, we build an auxiliary model comprising multiple stages where the layer weights in each stage are shared, after which we train it through distillation. Subsequently, we expand these learngene layers containing stage information at their corresponding stage to initialize models of variable depths. Extensive experiments on ImageNet-1K demonstrate that SWS achieves consistent better performance compared to many models trained from scratch, while reducing around 6.6x total training costs. In some cases, SWS performs better only after 1 epoch tuning. When initializing variable-sized models adapting for different resource constraints, SWS achieves better results while reducing around 20x parameters stored to initialize these models and around 10x pre-training costs, in contrast to the pre-training and fine-tuning approach.

Allowing the generation of effective interactions between distant quantum emitters (QEs) via flying photons, quantum interconnect (QI) is essentially a light-matter interface and acts as a building block in quantum technologies. A surface plasmon polariton (SPP) supported by a metallic waveguide provides an ideal interface to explore strong light-matter couplings and to realize QI. However, the loss of SPP in metal makes the mediated entanglement of the QEs damp with the increase of the distance and time, which hinders its applications. We propose a scheme of non-Markovian QI formed by the SPP of a metallic nanowire. A mechanism to make the generated entanglement of the QEs persistent is discovered. We find that, as long as bound states are formed in the energy spectrum of total QE-SPP system, the damping of the SPP-mediated entanglement is overcome even in the presence of the metal absorption to the SPP. Our finding enriches our understanding of light-matter couplings in absorptive medium and paves the way for using the SPP in designing QI.

In this work, we perform a lattice QCD study of the intrinsic, rapidity-independent soft function within the framework of large momentum effective theory. The computation is carried out using a gauge ensemble of $N_f=2+1+1$ clover-improved twisted mass fermion. After applying an appropriate renormalization procedure and the removal of significant higher-twist contamination, we obtain the intrinsic soft function that is comparable to the one-loop perturbative result at large external momentum. The determination of the nonperturbative soft function from first principles is crucial to sharpen our understanding of the processes with small transverse momentum such as the Drell-Yan production and the semi-inclusive deep inelastic scattering. Additionally, we calculate the Collins-Soper evolution kernel using the quasi-transverse-momentum-dependent wave function as input., Comment: 7 pages, 4 figures

Two field-sparsening methods, namely the sparse-grid method and the random field selection method, are used in this paper for the construction of the 2-point and 3-point correlation functions in lattice QCD. We argue that, due to the high correlation among the lattice correlators at different field points associated with source, current, and sink locations, one can save a lot of computational time by performing the summation over a subset of the lattice sites. Furthermore, with this strategy, one only needs to store a small fraction of the full quark propagators. It is found that the number of field points can be reduced by a factor of $\sim$100 for the point-source operator and a factor of $\sim$1000 for the Gaussian-smeared operator, while the uncertainties of the correlators only increase by $\sim$15\%. Therefore, with a modest cost of the computational resources, one can approach the precision of the all-to-all correlators using the field-sparsening methods., Comment: 16 pages, 9 figures

We present an exploratory lattice QCD calculation of the neutrinoless double beta decay $\pi\pi\to ee$. Under the mechanism of light-neutrino exchange, the decay amplitude involves significant long-distance contributions. The calculation reported here, with pion masses $m_\pi=420$ and 140 MeV, demonstrates that the decay amplitude can be computed from first principles using lattice methods. At unphysical and physical pion masses, we obtain that amplitudes are $24\%$ and $9\%$ smaller than the predication from leading order chiral perturbation theory. Our findings provide the lattice QCD inputs and constraints for effective field theory. A follow-on calculation with fully controlled systematic errors will be possible with adequate computational resources., Comment: 6 pages, 3 figures. V2: version accepted by PRL; minor changes compared to v1