Archives AI News

Correcting Auto-Differentiation in Neural-ODE Training

arXiv:2306.02192v2 Announce Type: replace Abstract: Does the use of auto-differentiation yield reasonable updates for deep neural networks (DNNs)? Specifically, when DNNs are designed to adhere to neural ODE architectures, can we trust the gradients provided by auto-differentiation? Through mathematical analysis and numerical evidence, we demonstrate that when neural networks employ high-order methods, such as Linear Multistep Methods (LMM) or Explicit Runge-Kutta Methods (ERK), to approximate the underlying ODE flows, brute-force auto-differentiation often introduces artificial oscillations in the gradients that prevent convergence. In the case of Leapfrog and 2-stage ERK, we propose simple post-processing techniques that effectively eliminates these oscillations, correct the gradient computation and thus returns the accurate updates.

Beyond Synthetic Augmentation: Group-Aware Threshold Calibration for Robust Balanced Accuracy in Imbalanced Learning

arXiv:2509.02592v1 Announce Type: new Abstract: Class imbalance remains a fundamental challenge in machine learning, with traditional solutions often creating as many problems as they solve. We demonstrate that group-aware threshold calibration--setting different decision thresholds for different demographic groups--provides superior robustness compared to synthetic data generation methods. Through extensive experiments, we show that group-specific thresholds achieve 1.5-4% higher balanced accuracy than SMOTE and CT-GAN augmented models while improving worst-group balanced accuracy. Unlike single-threshold approaches that apply one cutoff across all groups, our group-aware method optimizes the Pareto frontier between balanced accuracy and worst-group balanced accuracy, enabling fine-grained control over group-level performance. Critically, we find that applying group thresholds to synthetically augmented data yields minimal additional benefit, suggesting these approaches are fundamentally redundant. Our results span seven model families including linear, tree-based, instance-based, and boosting methods, confirming that group-aware threshold calibration offers a simpler, more interpretable, and more effective solution to class imbalance.

Preference Robustness for DPO with Applications to Public Health

arXiv:2509.02709v1 Announce Type: new Abstract: We study an LLM fine-tuning task for designing reward functions for sequential resource allocation problems in public health, guided by human preferences expressed in natural language. This setting presents a challenging testbed for alignment due to complex and ambiguous objectives and limited data availability. We propose DPO-PRO, a robust fine-tuning algorithm based on Direct Preference Optimization (DPO), which accounts for uncertainty in the preference distribution using a lightweight Distributionally Robust Optimization (DRO) formulation. Unlike prior DRO-based DPO methods, DPO-PRO is significantly less conservative. We evaluate DPO-PRO on a real-world maternal mobile health program operated by the non-profit organization ARMMAN, as well as on standard alignment benchmarks. Experimental results demonstrate that our method consistently improves robustness to noisy preference signals compared to existing DPO variants. Moreover, DPO-PRO achieves comparable performance to prior self-reflection-based baseline for reward function design, while requiring significantly lower inference-time cost.

Imitate Optimal Policy: Prevail and Induce Action Collapse in Policy Gradient

arXiv:2509.02737v1 Announce Type: new Abstract: Policy gradient (PG) methods in reinforcement learning frequently utilize deep neural networks (DNNs) to learn a shared backbone of feature representations used to compute likelihoods in an action selection layer. Numerous studies have been conducted on the convergence and global optima of policy networks, but few have analyzed representational structures of those underlying networks. While training an optimal policy DNN, we observed that under certain constraints, a gentle structure resembling neural collapse, which we refer to as Action Collapse (AC), emerges. This suggests that 1) the state-action activations (i.e. last-layer features) sharing the same optimal actions collapse towards those optimal actions respective mean activations; 2) the variability of activations sharing the same optimal actions converges to zero; 3) the weights of action selection layer and the mean activations collapse to a simplex equiangular tight frame (ETF). Our early work showed those aforementioned constraints to be necessary for these observations. Since the collapsed ETF of optimal policy DNNs maximally separates the pair-wise angles of all actions in the state-action space, we naturally raise a question: can we learn an optimal policy using an ETF structure as a (fixed) target configuration in the action selection layer? Our analytical proof shows that learning activations with a fixed ETF as action selection layer naturally leads to the AC. We thus propose the Action Collapse Policy Gradient (ACPG) method, which accordingly affixes a synthetic ETF as our action selection layer. ACPG induces the policy DNN to produce such an ideal configuration in the action selection layer while remaining optimal. Our experiments across various OpenAI Gym environments demonstrate that our technique can be integrated into any discrete PG methods and lead to favorable reward improvements more quickly and robustly.

LExI: Layer-Adaptive Active Experts for Efficient MoE Model Inference

arXiv:2509.02753v1 Announce Type: new Abstract: Mixture-of-Experts (MoE) models scale efficiently by activating only a subset of experts per token, offering a computationally sparse alternative to dense architectures. While prior post-training optimizations, such as inter- and intra-expert pruning, reduce memory usage they provide limited gains in inference-time compute efficiency. Moreover, existing MoE architectures typically activate a fixed number of experts uniformly across all layers, resulting in redundant computation and suboptimal performance. In this work, we first demonstrate that MoE pruning strategies improve only the memory footprint but do not significantly improve inference performance on GPU using optimized frameworks such as vLLM. To address this, we introduce LExI, a data-free optimization technique that determines the optimal number of active experts per layer in a pretrained MoE model. LExI leverages only the model weights to estimate the relative importance of each layer and adaptively assigns the number of active experts accordingly per layer. Experiments on state-of-the-art language and vision MoE benchmarks demonstrate that LExI significantly outperforms traditional MoE pruning approaches in terms of inference efficiency with negligible accuracy loss. For example, using LExI, Qwen1.5-MoE achieves the same throughput on Nvidia H100 GPU with 10% better accuracy than traditional expert pruning.

Mentality: A Mamba-based Approach towards Foundation Models for EEG

arXiv:2509.02746v1 Announce Type: new Abstract: This work explores the potential of foundation models, specifically a Mamba-based selective state space model, for enhancing EEG analysis in neurological disorder diagnosis. EEG, crucial for diagnosing conditions like epilepsy, presents significant challenges due to its noisy, high-dimensional, and nonlinear nature. Traditional machine learning methods have made advances in automating EEG analysis but often fail to capture its complex spatio-temporal dynamics. Recent advances in deep learning, particularly in sequence modeling, offer new avenues for creating more generalized and expressive models capable of handling such complexities. By training a Mamba-based model on a large dataset containing seizure and non-seizure EEG recordings through a self-supervised reconstruction task followed by a seizure detection task, we demonstrate the model's effectiveness, achieving an AUROC of 0.72 on a held-out test set. This approach marks a significant step toward developing large-scale, clinically applicable foundation models for EEG data analysis.

FlowKac: An Efficient Neural Fokker-Planck solver using Temporal Normalizing Flows and the Feynman-Kac Formula

arXiv:2503.11427v2 Announce Type: replace Abstract: Solving the Fokker-Planck equation for high-dimensional complex dynamical systems remains a pivotal yet challenging task due to the intractability of analytical solutions and the limitations of traditional numerical methods. In this work, we present FlowKac, a novel approach that reformulates the Fokker-Planck equation using the Feynman-Kac formula, allowing to query the solution at a given point via the expected values of stochastic paths. A key innovation of FlowKac lies in its adaptive stochastic sampling scheme which significantly reduces the computational complexity while maintaining high accuracy. This sampling technique, coupled with a time-indexed normalizing flow, designed for capturing time-evolving probability densities, enables robust sampling of collocation points, resulting in a flexible and mesh-free solver. This formulation mitigates the curse of dimensionality and enhances computational efficiency and accuracy, which is particularly crucial for applications that inherently require dimensions beyond the conventional three. We validate the robustness and scalability of our method through various experiments on a range of stochastic differential equations, demonstrating significant improvements over existing techniques.

The Transparent Earth: A Multimodal Foundation Model for the Earth’s Subsurface

arXiv:2509.02783v1 Announce Type: new Abstract: We present the Transparent Earth, a transformer-based architecture for reconstructing subsurface properties from heterogeneous datasets that vary in sparsity, resolution, and modality, where each modality represents a distinct type of observation (e.g., stress angle, mantle temperature, tectonic plate type). The model incorporates positional encodings of observations together with modality encodings, derived from a text embedding model applied to a description of each modality. This design enables the model to scale to an arbitrary number of modalities, making it straightforward to add new ones not considered in the initial design. We currently include eight modalities spanning directional angles, categorical classes, and continuous properties such as temperature and thickness. These capabilities support in-context learning, enabling the model to generate predictions either with no inputs or with an arbitrary number of additional observations from any subset of modalities. On validation data, this reduces errors in predicting stress angle by more than a factor of three. The proposed architecture is scalable and demonstrates improved performance with increased parameters. Together, these advances make the Transparent Earth an initial foundation model for the Earth's subsurface that ultimately aims to predict any subsurface property anywhere on Earth.

Learn and Unlearn: Addressing Misinformation in Multilingual LLMs

arXiv:2406.13748v3 Announce Type: replace-cross Abstract: This paper investigates the propagation of harmful information in multilingual large language models (LLMs) and evaluates the efficacy of various unlearning methods. We demonstrate that fake information, regardless of the language it is in, once introduced into these models through training data, can spread across different languages, compromising the integrity and reliability of the generated content. Our findings reveal that standard unlearning techniques, which typically focus on English data, are insufficient in mitigating the spread of harmful content in multilingual contexts and could inadvertently reinforce harmful content across languages. We show that only by addressing harmful responses in both English and the original language of the harmful data can we effectively eliminate generations for all languages. This underscores the critical need for comprehensive unlearning strategies that consider the multilingual nature of modern LLMs to enhance their safety and reliability across diverse linguistic landscapes.

Neural Canonical Polyadic Factorization for Traffic Analysis

arXiv:2506.15079v4 Announce Type: replace Abstract: Modern intelligent transportation systems rely on accurate spatiotemporal traffic analysis to optimize urban mobility and infrastructure resilience. However, pervasive missing data caused by sensor failures and heterogeneous sensing gaps fundamentally hinders reliable traffic modeling. This paper proposes a Neural Canonical Polyadic Factorization (NCPF) model that synergizes low-rank tensor algebra with deep representation learning for robust traffic data imputation. The model innovatively embeds CP decomposition into neural architecture through learnable embedding projections, where sparse traffic tensors are encoded into dense latent factors across road segments, time intervals, and mobility metrics. A hierarchical feature fusion mechanism employs Hadamard products to explicitly model multilinear interactions, while stacked multilayer perceptron layers nonlinearly refine these representations to capture complex spatiotemporal couplings. Extensive evaluations on six urban traffic datasets demonstrate NCPF's superiority over six state-of-the-art baselines. By unifying CP decomposition's interpretable factor analysis with neural network's nonlinear expressive power, NCPF provides a principled yet flexible approaches for high-dimensional traffic data imputation, offering critical support for next-generation transportation digital twins and adaptive traffic control systems.