Stories tagged “Machine Learning.”
30 stories
sktime is a Python library offering a unified interface for various time series machine learning tasks, including forecasting, classification, and clustering. It provides dedicated algorithms and scikit-learn compatible tools for model building and validation.
Researchers introduce BI-Agent and BI-Bench to automate end-to-end business intelligence tasks, finding even advanced LLMs struggle with traditional BI workflows.
AutoGluon automates machine learning for tabular, time series, and multimodal data, finding the best model combination with minimal code.
Physics-informed neural networks (PINNs) may exhibit a failure mode where function values are accurate but derivatives are not, a problem highlighted by new benchmark evidence.
The paper introduces BudgetBench, a new protocol and reference harness for evaluating memory strategies in local large language model (LLM) agents by treating the per-call input-token budget as the independent variable. It aims to expose performance nuances that fixed-bud…
This paper introduces Physics-Informed Conformal Prediction (PI-CP), a framework that provides rigorous, spatially adaptive uncertainty estimates for neural operators solving partial differential equations (PDEs). It ensures provable coverage guarantees by embedding PDE r…
This paper introduces Fundamental Dynamical Units (FDUs) to infer interaction structures in complex networked systems from perturbation time-series data, addressing challenges like combinatorial complexity and causal ambiguity.
PyTorch is a Python package offering GPU-accelerated tensor computation akin to NumPy and a deep neural network framework built on a dynamic, tape-based autograd system. It allows seamless integration with existing Python libraries for flexible and fast deep learning deve…
Researchers have developed M3-Former, a multimodal transformer framework that uses large language models and a Mixture-of-Experts architecture to improve long-term vessel trajectory prediction.
A new paper introduces "Halo," a modification to deep forecasters that uses heteroscedastic estimation to improve point forecast accuracy, challenging previous findings. It adapts existing models to estimate a scale parameter alongside a location parameter, demonstrating …
A new benchmark, AhaBench, evaluates whether language agents truly learn from prior experience in long-horizon tasks, moving beyond single-prompt evaluations. It assesses agents' ability to improve behavior after receiving useful experience, even when direct support is re…
A "Fourier auxiliary head" is proposed to combat "physical representation laziness" in latent world models. It improves planning success and data efficiency by enforcing physically-informed latent space structuring.
ProToMEx is a new model-agnostic framework that uses Probabilistic Topic Models (PTMs) to provide rapid, interpretable explanations for machine learning classifiers, moving beyond simple feature attribution to reveal complex, combinatorial patterns. It offers both global …
Researchers have developed a data fusion framework that enhances the predictive accuracy of aerospace surrogate models by integrating experimental wind-tunnel observations with computational fluid dynamics (CFD) simulations.
A new paper introduces Q-MET, a quantum-assisted framework for Wi-Fi-based human activity recognition (HAR) that significantly reduces memory consumption during training and inference. It uses a hybrid quantum-classical neural network and structured pruning to achieve hig…
This study evaluates large language models (LLMs) for predicting weather-related forced outages in electricity distribution grids using a zero-shot approach. It benchmarks four LLMs against two supervised machine learning models, finding that while supervised models curre…
The BER-PEF framework offers a unified method for evaluating human mobility predictability by estimating the Bayes Error Rate, addressing the challenge of unobservable ground truth in real-world data. It provides a consistent protocol for comparing different prediction es…
New method for discovering unknown nonlinear dynamics from single state trajectory data using functional analysis and operator theory.
Researchers introduce "equation recast," a method that transforms parametric operator learning into a single canonical operator by analytically incorporating parameter variations. This enables zero-shot prediction across new parameter regimes and enhances data efficiency …
A new study reveals that large language models possess a "direction of ignorance" within their architecture, allowing them to dynamically adjust their reliance on pre-existing knowledge (Bayesian priors) as more contextual information becomes available.
A new framework called WMLLM introduces self-evolving AI optimization agents that use a "predict-then-act" world modeling approach to tackle complex black-box optimization problems. It leverages large language models to predict promising directions, significantly enhancin…
New method for aligning latent spaces without needing paired data.
A new method called MCC-PGPSE enhances parallel reinforcement learning by assigning "marginal coverage credit" to individual policies. This approach reduces redundant exploration and promotes diverse state-space coverage, leading to improved learning efficiency in multi-a…
Post-training quantization, an optimization for deploying LLMs, can inadvertently create a "validation--deployment gap" where models pass initial checks but exhibit malicious behavior after compression. This research formalizes this gap and demonstrates how latent backdoo…
Candle is a Rust-based ML framework from Hugging Face, prioritizing performance and ease of use for AI development.
Proposes a new architecture for property graph learning that integrates a small language model into graph message selection.
A new fuzzing framework, NeuronFuzz, is introduced to improve safety evaluation of Large Language Models (LLMs) by using internal safety neurons.
Researchers developed Dynamic Influence-Weighted Distillation (DIW) to improve single-IMU activity recognition. DIW uses multi-IMU data during training to enhance a model that only uses one IMU for inference, achieving significant performance gains.
Researchers have introduced GreenLeaf Law Embed Tiny, a 0.6 billion parameter embedding model designed for legal domain retrieval, achieving competitive performance on key benchmarks.
A new research paper introduces Equivariant Cellular Sheaf Networks, a framework that models molecular single-particle Hamiltonians as Laplacians of cellular sheaves, generalizing existing E(3)-equivariant methods.