Search for: All records

Deep learning models are widely used in decision-making and recommendation systems, where they typically rely on the assumption of a static data distribution between training and deployment. However, real-world deployment environments often violate this assumption. Users who receive negative outcomes may adapt their features to meet model criteria, i.e., recourse action. These adaptive behaviors create shifts in the data distribution and when models are retrained on this shifted data, a feedback loop emerges: user behavior influences the model, and the updated model in turn reshapes future user behavior. Despite its importance, this bidirectional interaction between users and models has received limited attention. In this work, we develop a general framework to model user strategic behaviors and their interactions with decision-making systems under resource constraints and competitive dynamics. Both the theoretical and empirical analyses show that user recourse behavior tends to push logistic and MLP models toward increasingly higher decision standards, resulting in higher recourse costs and less reliable recourse actions over time. To mitigate these challenges, we propose two methods—Fair-top-k and Dynamic Continual Learning (DCL)—which significantly reduce recourse cost and improve model robustness. Our findings draw connections to economic theories, highlighting how algorithmic decision-making can unintentionally reinforce a higher standard and generate endogenous barriers to entry. 

ABSTRACT Plant cytokinesis results in the formation of the cell plate by the phragmoplast which contains dynamic microtubules serving as the track for the delivery of cell wall builders included in Golgi vesicles. During the centrifugal process of cell plate assembly, new microtubules are assembled and bundled at the leading edge to prepare for vesicle transport while older microtubules are disassembled at the lagging edge upon the completion of vesicle delivery. The turnover of phragmoplast microtubules in this process is thought to be regulated by phosphorylation of the key microtubule bundling factor MAP65. A recent study revealed a surprising role of theα‐Aurora kinase, which is typically known for its role in governing the formation of the bipolar spindle apparatus, in phosphorylating the primary microtubule bundler MAP65‐3 in Arabidopsis. This phosphorylation positively contributes to the expansion of the phragmoplast. The phragmoplast midzone is also the hub for other cytokinesis‐important kinases. It is intriguing how these kinases are targeted and how they may crosstalk with each other to orchestrate the expansion of the phragmoplast. 

Abstract The evolutionarily conserved microspherule protein 1 (MCRS1) has diverse functions, ranging from transcriptional regulation to stabilization of microtubule minus ends in acentrosomal spindles in mammals. A previous study suggested that in the model plant Arabidopsis thaliana, inactivation of an MCRS1 homolog gene led to aborted embryogenesis. To test whether this lethality was caused solely by sporophytic defects, we used the heterozygous emb1967-1/mcrs1-1 mutant for reciprocal crosses with the wild-type plant and found that the MCRS1 gene was dispensable for the development of both male and female gametophytes. An MCRS1–GFP fusion protein was expressed in the mcrs1 mutant and suppressed the mutation as evidenced by restored growth. This functional fusion protein exclusively localized to interphase nuclei and became unnoticeable during mitosis before reappearing in the reforming daughter nuclei. Affinity purification of the MCRS1–GFP protein specifically recovered the Myb-like transcription factor DRMY1 (Development Regulated Myb-like 1) but not microtubule-associated factors. Direct MCRS1–DRMY1 interaction was also demonstrated by a localization-based assay in living cells. Thus, we hypothesized that MCRS1’s function was perhaps linked to transcription factors like DRMY1 and its paralog DP1 for regulation of gene expression during sporophyte development. 

ABSTRACT Actin microfilaments (F-actin) serve as the track for directional movement of organelles in plant cells. In actively growing plant cells, F-actin often form robust bundles that trespass the cellular dimension. To test how the F-actin network was employed for peroxisome movement, we wished to disturb actin organization by genetically compromising the function of villin (VLN) proteins that serve as the primary bundling factor inArabidopsis thalianacells. To do so, we isolated T-DNA insertional mutants in threeVLNgenes that were most actively expressed in vegetative tissues. We found that thevln4mutation greatly enhanced the growth defects caused by thevln2 vln3double mutant as thevln2 vln3 vln4triple mutant had a great reduction of organ growth and formed heavily deformed tissues. Both VLN2 and VLN4 proteins were detected on bundled F-actin filaments. Compared to the wild-type cells, the double and triple mutants exhibited progressively reduction of stable F-actin bundles and had fine F-actin filaments undergo rapid remodeling. The defective F-actin network did not prevent peroxisomes from taking on both rapid and slow movements along the F-actin tracks. However, we found that compromised F-actin bundling caused significant reductions in the speed of peroxisome movement and the displacement distance of peroxisome positions. Using a correlation analysis method, we also demonstrated that the complex heterogeneous peroxisome movement may be classified into clusters reflecting the directionality of peroxisome movement. The triple mutant suffered from a significant reduction of peroxisomes exhibiting long-range and linear movement. Our results provided insights into how VLN-dependent F-actin organization was coupled with the complex patterns of peroxisome movement. 

Abstract Spindle assembly in vertebrates requires the Aurora kinase, which is targeted to microtubules and activated by TPX2 (Targeting Protein of XKLP2). In Arabidopsis (Arabidopsis thaliana), TPX2-LIKE 3 (TPXL3), but not the highly conserved TPX2, is essential. To test the hypothesis that TPXL3 regulates the function of α Aurora kinase in spindle assembly, we generated transgenic Arabidopsis lines expressing an artificial microRNA targeting TPXL3 mRNA (amiR-TPXL3). The resulting mutants exhibited growth retardation, which was linked to compromised TPXL3 expression. In the mutant cells, α Aurora was delocalized from spindle microtubules to the cytoplasm, and spindles were assembled without recognizable poles. A functional TPXL3-GFP fusion protein first prominently appeared on the prophase nuclear envelope. Then, TPXL3-GFP localized to spindle microtubules (primarily toward the spindle poles, like γ-tubulin), and finally to the re-forming nuclear envelope during telophase and cytokinesis. However, TPXL3 was absent from phragmoplast microtubules. In addition, we found that the TPXL3 N-terminal Aurora-binding motif, microtubule-binding domain, and importin-binding motif, but not the C-terminal segment, were required for its mitotic function. Expression of truncated TPXL3 variants enhanced the defects in spindle assembly and seedling growth of amiR-TPXL3 plants. Taken together, our findings uncovered the essential function of TPXL3, but not TPX2, in targeting and activating α Aurora kinase for spindle apparatus assembly in Arabidopsis. 

The Arctic Oscillation (AO) has been observed to undergo distinct decadal structural fluctuations that significantly influence regional weather and climate. Understanding the drivers and mechanisms behind the AO’s spatial nonstationarity is critical for improving climate predictions related to the AO. Wepresent evidence that the Atlantic Multidecadal Oscillation (AMO) plays a pivotal role in modulating AO’s Pacific center in recent decades. The poleward amplified cooling associated with negative AMO enhances the north-south temperature gradient and results the strengthened westerly winds and stratospheric polar vortex (SPV) responses, which reflects more planetary waves from the North Pacific to the North Atlantic. This enhances the atmospheric coupling between these regions and leads to amore pronounced Pacific center within theAOpattern.Numerical simulations fromECHAM5 and 35 CMIP6 models further corroborate the essential role of the AMO. These findings advance our understanding of the mechanisms driving the variability of the AO pattern. 

Nanohybrids of graphene and colloidal semiconductor quantum dots (QDs/Gr) provide a promising quantum sensing scheme for photodetection. Despite exciting progress made in QDs/Gr photodetectors in broadband from ultraviolet to short-wave infrared, the device performance is limited in middle-wave infrared (MWIR) detection. A fundamental question arises as to whether the thermal noiseinduced dark current and hence poor signal-to-noise ratio in conventional uncooled MWIR photodetectors persist in QDs/ Gr nanohybrids. Herein, we investigated noise, responsivity (R*), and specific detectivity (D*) in HgTe QDs/Gr nanohybrids, revealing that the noise and R* are decoupled in nanohybrids and each can be optimized independently toward its theoretical limit. Specifically, the noise in the QDs/Gr nanohybrids is dominated by that of graphene with a negligible effect from the dark current in HgTe QDs and can be optimized to its intrinsic limit by removing charge doping of adsorbed polar molecules on graphene. Furthermore, the R* is proportional to the photoconductive gain enabled by the strong quantum confinement in QDs and Gr. Achieving high gain in the MWIR spectrum, however, is challenging and requires elimination of charge traps primarily from the surface states of the narrow-bandgap semiconductor HgTe QDs. Using grain-rotation-induced grain-coalescence growth of single-layer and core/shell HgTe QDs, we show the that HgTe QDs surface states caused by Te deficiency can be dramatically suppressed, resulting in high gain up to 4.0 × 107 in the MWIR spectrum. The optimized noise and R* have led to high uncooled MWIR D* up to 2.4 × 1011 Jones, making nanohybrids promising to surpass the fundamental dark-current limit in conventional photodetectors. 

Topology has emerged as a field for describing and controlling order and matter, and thereby the physical properties of materials. There are several largely disparate fields focused on examining and manipulating topology. One of these arenas is in the realm of real space, manipulating systems in terms of their spatial properties, to control the corresponding structural, mechanical, and self- assembling responses. Much of the work in soft matter topology falls within this domain. A second arena is in the domain of momentum or k-space wherein topology controls the features of the electronic band structure of materials, and topologically non-trivial features result in the development of materials with truly unique properties. This work focuses squarely on the realm of condensed matter physics. Here, we review concepts of real- and k-space topology and propose areas for convergence between these two disparate fields. 

Recently, a wide range of memory-efficient LLM training algorithms have gained substantial popularity. These methods leverage the low-rank structure of gradients to project optimizer states into a subspace using a projection matrix found by singular value decomposition (SVD). However, convergence of these algorithms is highly dependent on the update rules of their projection matrix. This work provides the first convergence guarantee for arbitrary update rules of projection matrices, generally applicable to optimizers that can be analyzed with Hamiltonian Descent, including common ones such as LION and Adam. Inspired by this theoretical understanding, the authors propose Online Subspace Descent, a new family of subspace descent optimizers that do not rely on SVD. Instead of updating the projection matrix with eigenvectors, Online Subspace Descent updates it with online PCA. This approach is flexible and introduces minimal overhead to training. Experiments show that for pretraining LLaMA models ranging from 60M to 7B parameters on the C4 dataset, Online Subspace Descent achieves lower perplexity and better downstream task performance than state-of-the-art low-rank training methods across settings, narrowing the gap with full-rank baselines.