Search for: All records

Contrastive learning (CL) has been widely investigated with various learning mech- anisms and achieves strong capability in learning representations of data in a self-supervised manner using unlabeled data. A common fashion of contrastive learning on this line is employing large-sized encoders to achieve comparable performance as the supervised learning counterpart. Despite the success of the labelless training, current contrastive learning algorithms failed to achieve good performance with lightweight (compact) models, e.g., MobileNet, while the re- quirements of the heavy encoders impede the energy-efficient computation, espe- cially for resource-constrained AI applications. Motivated by this, we propose a new self-supervised CL scheme, named SACL-XD, consisting of two technical components, Slimmed Asymmetrical Contrastive Learning (SACL) and Cross- Distillation (XD), which collectively enable efficient CL with compact models. While relevant prior works employed a strong pre-trained model as the teacher of unsupervised knowledge distillation to a lightweight encoder, our proposed method trains CL models from scratch and outperforms them even without such an expensive requirement. Compared to the SoTA lightweight CL training (dis- tillation) algorithms, SACL-XD achieves 1.79% ImageNet-1K accuracy improve- ment on MobileNet-V3 with 64⇥ training FLOPs reduction. Code is available at https://github.com/mengjian0502/SACL-XD. 

CsYbSe2 has an ideal triangular-lattice geometry with pronounced two-dimensionality, pseudospin-1/2 nature, and the absence of structural disorder. These excellent characteristics favor a quantum spin-liquid realization in this material. In this work, we applied quasihydrostatic compression methods to explore the structural behaviors. Our study reveals that CsYbSe2 undergoes a structural transition around 24 GPa, accompanied by a large volume collapse of ΔV /V0∼13%. The ambient hexagonal structure with the space group P63/mmcis lowered to the tetragonal structure (P4/mmm) under high pressure. Meanwhile, the color of CsYbSe2 changes gradually from red to black before the transition. Dramatic pressure-induced changes are clarified by the electronic structure calculations from the first principles, which indicate that the initial insulating ground state turns metallic in a squeezed lattice. These findings highlight Yb-based dichalcogenide delafossites as an intriguing material to probe novel quantum effects under high pressure. 

Abstract The absorption by neutral hydrogen in the intergalactic medium (IGM) produces the Ly α forest in the spectra of quasars. The Ly α forest absorbers have a broad distribution of neutral hydrogen column density N H I and Doppler b parameter. The narrowest Ly α absorption lines (of lowest b ) with neutral hydrogen column density above ∼10 13 cm −2 are dominated by thermal broadening, which can be used to constrain the thermal state of the IGM. Here we constrain the temperature-density relation T = T 0 ( ρ / ρ ¯ ) γ − 1 of the IGM at 1.6 < z < 3.6 by using N H I and b parameters measured from 24 high-resolution and high-signal-to-noise quasar spectra and by employing an analytic model to model the N H I -dependent low- b cutoff in the b distribution. In each N H I bin, the b cutoff is estimated using two methods, one non-parametric method from computing the cumulative b distribution and a parametric method from fitting the full b distribution. We find that the IGM temperature T 0 at the mean gas density ρ ¯ shows a peak of ∼1.5 × 10 4 K at z ∼ 2.7–2.9. At redshift higher than this, the index γ approximately remains constant, and it starts to increase toward lower redshifts. The evolution in both parameters is in good agreement with constraints from completely different approaches, which signals that He ii reionization completes around z ∼ 3. 

ABSTRACT De novo root regeneration (DNRR) is a developmental process that regenerates adventitious roots from wounded tissues. Phytohormone signaling pathways involved in microbial resistance are mobilized after cutting and influence de novo root regeneration. Microbes may positively or negatively influence the development and stress responses of a plant. However, most studies on the molecular mechanisms of de novo organogenesis are performed in aseptic conditions. Thus, the potential crosstalk between organ regeneration and biotic stresses is underexplored. Here, we report the development of a versatile experimental system to study the impact of microbes on DNRR. Using this system, we found that bacteria inhibited root regeneration by activation of, but not limited to, pathogen-associated molecular pattern (PAMP)-triggered immunity. Sensing bacteria-derived flagellin 22 peptide (flg22) inhibited root regeneration by interfering with the formation of an auxin maximum at the wound site. This inhibition relies on the receptor complex that recognizes microbial patterns but may bypass the requirement of salicylic acid signaling. 

Moiré potential profile can form flat electronic bands and manifest correlated states of electrons, where carrier doping is essential for observing those correlations. In this work, we uncover a hidden but remarkable many-electron effect: doped carriers form a two-dimensional plasmon and strongly couple with quasiparticles to renormalize moiré potential and realize ultra-flat bands. Using many-body perturbation theory, we demonstrate this effect in twisted MoS₂/WS₂heterobilayer. The moiré potential is significantly enhanced upon carrier doping, and the bandwidth is reduced by order of magnitude, leading to drastic quenching of electronic kinetic energy and stronger correlation. We further predict that the competition between correlated mechanisms can be effectively controlled via doping, giving hope to a quantum transition between Mott and charge-transfer insulating states. Our work reveals that the potential renormalization effect of doping is much more significant in determining and controlling many-electron electronic correlations than sole filling-factor tuning in semiconducting moiré crystals.

 

Moiré superlattices host a rich variety of correlated electronic phases. However, the moiré potential is fixed by interlayer coupling, and it is dependent on the nature of carriers and valleys. In contrast, it has been predicted that twisted hexagonal boron nitride (hBN) layers can impose a periodic electrostatic potential capable of engineering the properties of adjacent functional layers. Here, we show that this potential is described by a theory of electric polarization originating from the interfacial charge redistribution, validated by its dependence on supercell sizes and distance from the twisted interfaces. This enables controllability of the potential depth and profile by controlling the twist angles between the two interfaces. Employing this approach, we further demonstrate how the electrostatic potential from a twisted hBN substrate impedes exciton diffusion in semiconductor monolayers, suggesting opportunities for engineering the properties of adjacent functional layers using the surface potential of a twisted hBN substrate. 

By learning a sequence of tasks continually, an agent in continual learning (CL) can improve the learning performance of both a new task and `old' tasks by leveraging the forward knowledge transfer and the backward knowledge transfer, respectively. However, most existing CL methods focus on addressing catastrophic forgetting in neural networks by minimizing the modification of the learnt model for old tasks. This inevitably limits the backward knowledge transfer from the new task to the old tasks, because judicious model updates could possibly improve the learning performance of the old tasks as well. To tackle this problem, we first theoretically analyze the conditions under which updating the learnt model of old tasks could be beneficial for CL and also lead to backward knowledge transfer, based on the gradient projection onto the input subspaces of old tasks. Building on the theoretical analysis, we next develop a ContinUal learning method with Backward knowlEdge tRansfer (CUBER), for a fixed capacity neural network without data replay. In particular, CUBER first characterizes the task correlation to identify the positively correlated old tasks in a layer-wise manner, and then selectively modifies the learnt model of the old tasks when learning the new task. Experimental studies show that CUBER can even achieve positive backward knowledge transfer on several existing CL benchmarks for the first time without data replay, where the related baselines still suffer from catastrophic forgetting (negative backward knowledge transfer). The superior performance of CUBER on the backward knowledge transfer also leads to higher accuracy accordingly. 

Abstract Quantum many-body systems in one dimension (1D) exhibit some peculiar properties. In this article, we review some of our work on strongly interacting 1D spinor quantum gas. First, we discuss a generalized Bose–Fermi mapping that maps the charge degrees of freedom to a spinless Fermi gas and the spin degrees of freedom to a spin chain model. This also maps the strongly interacting system into a weakly interacting one, which is amenable for perturbative calculations. Next, based on this mapping, we construct an ansatz wavefunction for the strongly interacting system, using which many physical quantities can be conveniently calculated. We showcase the usage of this ansatz wavefunction by considering the collective excitations and quench dynamics of a harmonically trapped system.