This content explores the technique of federated learning, specifically focusing on how quantized global model updates can optimize its efficiency. It likely delves into methods for reducing communication overhead and computational costs in distributed machine learning environments.
Physics-Informed Neural Networks (PINNs) often suffer from slow convergence and instability due to complex loss landscapes. This paper proposes a lightweight, curvature-aware optimization framework that augments existing first-order optimizers to improve convergence speed, training stability, and solution accuracy on partial differential equations (PDEs).
This research introduces Transcriptome-based Drug Design (TBDD) as a generative inverse problem to design drug molecules conditioned on desired transcriptomic state transitions. It proposes "ACURE" (A Cellular Response Engine), a multi-resolution transcriptome-guided diffusion framework, to address the challenges of this complex task.
Este artigo apresenta o AgriPriceBD, um novo conjunto de dados diário de preços de commodities agrícolas de Bangladesh, extraído com auxílio de LLM. Ele avalia sete abordagens de previsão, incluindo modelos clássicos e arquiteturas de deep learning, para estabilização da renda e segurança alimentar.
This research discovers that LoRA fine-tuning leads to 'un-learning' on contested examples, where high annotator disagreement correlates with increased loss during training. This pattern is distinct from full fine-tuning and consistently observed across multiple encoder and decoder-only models and datasets.
When2Speak is a new synthetic dataset and four-stage generation pipeline designed to teach Large Language Models (LLMs) appropriate intervention timing in multi-party conversations. It addresses the challenge of avoiding excessive interruptions and improving conversational coherence in group interactions.
This manuscript introduces Data Driven Variational Basis Learning (DVBL), a novel non-neural framework for learning data-adaptive basis functions directly from high-dimensional data. It provides an explicit, interpretable, and mathematically transparent alternative to neural networks for representation learning, addressing their limitations in control and transparency.
This research introduces EasyRL, a novel data-efficient reinforcement learning approach for self-evolving LLMs, designed to overcome high annotation costs and performance issues in existing methods. Inspired by cognitive learning theory, EasyRL integrates knowledge transfer from easy labeled data with a progressive divide-and-conquer strategy for difficult unlabeled data.
This paper introduces a self-contained TTS-STT flywheel to close the gap in niche-domain Indic ASR where commercial and open-source systems fail. It synthesizes entity-dense audio to significantly improve the Entity-Hit-Rate on challenging datasets for languages like Telugu.
This paper introduces ANDRE, a novel Attention-based Neuro-symbolic Differentiable Rule Extractor (ILP) framework for learning first-order logic programs. It optimizes over a continuous rule space with fully differentiable, attention-driven logical operators, addressing scalability challenges in noisy and probabilistic settings.
CAWI proposes a new weight initialization framework for Randomized Neural Networks (RdNNs) that addresses the limitation of conventional random initialization ignoring inter-feature dependence. It uses a data-fitted copula to ensure frozen projections respect empirical dependence, improving conditioning and predictive performance.
OceanCBM is the first concept bottleneck model (CBM) for spatiotemporal prediction and mechanistic interrogation of ocean dynamics. It predicts mixed layer heat content, a precursor to marine heatwaves, using mixed supervision and prescribed geophysical fluid dynamics concepts to ensure fidelity to ground-truth physics.
This research benchmarks classical (Lasso, Ridge) and Bayesian (Horseshoe, Spike-and-Slab) sparse regression methods under challenging conditions like correlated features and weak signals. Bayesian methods generally outperform classical ones in prediction error, with Horseshoe offering excellent coverage, though classical methods are faster.
This paper introduces Adaptive Multi-Scale Goodness Aggregation (AMSGA), a novel extension of the Forward-Forward (FF) algorithm to enhance stability, robustness, and generalization in local-learning neural networks. AMSGA incorporates multi-scale goodness aggregation, adaptive hard negative mining, and layer-dependent thresholds. Experiments on MNIST and Fashion-MNIST show consistent performance improvements over the baseline FF algorithm.
The paper introduces Hierarchical Multi-view HAAR (HMH), a novel spectral graph-learning framework designed to address oversmoothing and degree-biased aggregation in GNNs for heterophilous graphs. HMH constructs a soft graph hierarchy and applies learnable spectral filters using sparse, orthonormal Haar bases, achieving near-linear time scalability.
Sign-Muon is a novel 1-bit, matrix-aware optimizer proposed for communication-efficient distributed training of large neural networks. It combines sign aggregation with a polar-step framework, achieving an O(1/√T) nonconvex rate and reducing the stochastic term by 1/√M for M workers.
This paper evaluates whether geospatial foundation model embeddings improve cross-country maize yield predictions in Sub-Saharan Africa. It finds that while within-country predictions are moderate, all feature sets, including foundation model embeddings, perform poorly under cross-country testing, indicating a significant generalisability gap.
This paper proposes a method for anomaly detection called Chimera Training, focusing on violations of semantic constraints given as logical rules over learned visual concepts. It employs a neural rule evaluator that compiles constraints into directed acyclic graphs, learning logical operators to calculate rule-satisfaction probabilities, even with scarce training data for actual violations.
This work introduces CroCo, a method for cross-lingual contrastive preference tuning on self-generated responses from LLMs, demonstrating effective transfer across 14 languages without language-specific preference annotations. An English-trained reward model yields useful rankings across most languages, improving existing models and preventing catastrophic forgetting, provided on-policy data is used.
This paper introduces Automatically Differentiable Nonlinear Tensor Networks (ADNTNs), a family of structured weight generators for exponential compression of Deep Neural Networks. It extends low-rank adaptation and tensor factorization by building large weight tensors through a hierarchy of small cores and nonlinear activations.