Post-training activation compression is essential for deploying Large Language Models (LLMs) on resource-constrained hardware.However, standard methods like Singular Value Decomposition (SVD) are gradient-blind: they preserve high-variance dimensions regardless of their impact on factual knowledge preservation.We introduce Fisher-Aligned Subspace Compression (FASC), a knowledge-aware compression framework that selects subspaces by directly modeling activation-gradient coupling, minimizing a second-order surrogate of the loss function.FASC leverages the Fisher Information Matrix to identify dimensions critical for factual knowledge, which often reside in low-variance but high-gradient-sensitivity subspaces.We propose the Dependence Violation Score (ρ) as a general-purpose diagnostic metric that quantifies activation-gradient coupling, revealing where factual knowledge is stored within transformer architectures.Extensive experiments on Mistral-7B and Llama-3-8B demonstrate that FASC preserves 6-8% more accuracy on knowledge-intensive benchmarks (MMLU, LAMA) compared to variance-based methods at 50% rank reduction, effectively enabling a 7B model to match the factual recall of a 13B uncompressed model.Our analysis reveals that ρ serves as a fundamental signal of stored knowledge, with high-ρ layers emerging only when models internalize factual associations during training.

Quantum machine learning for spin and molecular systems faces critical challenges of scarce labeled data and computationally expensive simulations. To address these limitations, we introduce Hamiltonian-Masked Autoencoding (HMAE), a novel self-supervised framework that pre-trains transformers on unlabeled quantum Hamiltonians, enabling efficient few-shot transfer learning. Unlike random masking approaches, HMAE employs a physics-informed strategy based on quantum information theory to selectively mask Hamiltonian terms based on their physical significance. Experiments on 12,500 quantum Hamiltonians (60% real-world, 40% synthetic) demonstrate that HMAE achieves 85.3% ± 1.5% accuracy in phase classification and 0.15 ± 0.02 eV MAE in ground state energy prediction with merely 10 labeled examples—a statistically significant improvement (p < 0.01) over classical graph neural networks (78.1% ± 2.1%) and quantum neural networks (76.8% ± 2.3%). Our method’s primary advantage is exceptional sample efficiency—reducing required labeled examples by 3-5× compared to baseline methods—though we emphasize that ground truth values for fine-tuning and evaluation still require exact diagonalization or tensor networks. We explicitly acknowledge that our current approach is limited to small quantum systems (specifically limited to 12 qubits during training, with limited extension to 16-20 qubits in testing) and that, while promising within this regime, this size restriction prevents immediate application to larger systems of practical interest in materials science and quantum chemistry.

Integrating large language models (LLMs) as action proposers in reinforcement learning (RL) boosts performance in text-based environments but incurs high computational costs.We introduce a cache-efficient framework for Bayesian RL with LLM-derived action suggestions, reducing costs while maintaining near-optimal performance.Our approach features a meta-learned adaptive cache, optimized via meta-learning based on policy performance, enabling efficient inference in text-based games (e.g., TextWorld, ALFWorld) and robotic control tasks (e.g., MuJoCo, Meta-World).It achieves a 3.8-4.7×reduction in LLM queries, 4.0-12.0×lower median latencies (85-93ms on consumer hardware), and retains 96-98% of uncached performance.Theoretical KL-divergence bounds ensure reliable cached decisions, validated empirically across tasks with 90.4-95.6%success rates in text environments.For offline RL, our CQL-Prior variant improves performance by 14-29% and reduces training time by 38-40%.Evaluations across eight diverse tasks demonstrate the framework's generalizability and practicality for resource-constrained settings, making LLMguided RL viable for text-based and robotic applications.

As AI deployment shifts to edge devices, efficient sequence modeling becomes critical.State-space models (SSMs), particularly Mamba, rival Transformers with linear-time complexity and strong performance across tasks, yet their large parameter counts hinder resource-constrained use.We propose a novel unstructured pruning framework tailored for Mamba, achieving up to 70% parameter reduction with only 3-9% performance loss.Unlike Transformer-focused pruning, our approach leverages Mamba's recurrent dynamics through: (1) pruning based on weight and gradient importance to preserve critical parameters, (2) a gradual pruning process to ensure model stability, and (3) a global strategy optimizing parameter allocation across the model.Extensive experiments on WikiText-103, Long Range Arena, and ETT benchmarks show significant efficiency gains, with 1.77× faster inference and 46% less memory.Our component analysis reveals Mamba's robustness, enabling practical deployment while requiring careful use to avoid biases in sensitive applications.

Crashes caused by distracted driving result in more than 3000 deaths every year in the U.S. Distracted driver behavior detection is instrumental for driver assist systems. Researchers have focused on autonomously detecting distracted driver behavior so that drivers can be alerted in time to reduce the risk of crashes. Despite the large number of approaches presented in the literature, there are still issues related to proper performance evaluation, reproducibility and lack of or very slow adoption of these approaches by the transportation industry. Most existing approaches either do not provide documented and usable codes or use private datasets, or do not present the experiment details, such as data split, sometimes resulting in inflated accuracy numbers. Moreover, these factors also make many results not reproducible. In addition, the performance metrics should be chosen carefully to measure various aspects of different methods, including their generalizability, and action localization ability in time. In this work, we perform a commensurate comparison of different state-of-the-art methods by using different data splits and performance metrics on the StateFarm distracted driving and AI CITY Challenge datasets. With the data split experiments, we highlight the importance of leave-N-driver-out cross validation, since these models should perform well in real-world testing with never-before-seen drivers. The results show the importance of data splitting and the performance metric for the comparison and evaluation of different methods, and their significant effects on the results.

Recognizing the activities causing distraction in real-world driving scenarios is critical for ensuring the safety and reliability of both drivers and pedestrians on the roadways. Conventional computer vision techniques are typically data-intensive and require a large volume of annotated training data to detect and classify various distracted driving behaviors, thereby limiting their generalization ability, efficiency and scalability. We aim to develop a generalized framework that showcases robust performance with access to limited or no annotated training data. Recently, vision-language models have offered large-scale visual-textual pretraining that can be adapted to task-specific learning like distracted driving activity recognition. Vision-language pretraining models like CLIP have shown significant promise in learning natural language-guided visual representations. This paper proposes a CLIP-based driver activity recognition approach that identifies driver distraction from naturalistic driving images and videos. CLIP’s vision embedding offers zero-shot transfer and task-based finetuning, which can classify distracted activities from naturalistic driving video. Our results show that this framework offers state-of-the-art performance on zero-shot transfer, finetuning and video-based models for predicting the driver’s state on four public datasets. We propose frame-based and video-based frameworks developed on top of the CLIP’s visual representation for distracted driving detection and classification tasks and report the results. Our code is available at https://github.com/zahid-isu/DriveCLIP

Driver behavior analysis plays an important role in driver assistance systems. A driver’s face and head pose hold the key towards understanding whether the driver’s attention and concentration are on the road while driving. Naturalistic driving studies (NDS) allow observing drivers in real-time under naturalistic traffic conditions. Yet, data collected in NDS often comprise low-resolution videos usually with more challenging camera positions compared to controlled studies. For instance, when the camera is not directly facing the driver, classifying head pose becomes more challenging, since the variation between different classes becomes much smaller. In this paper, we propose three different approaches to classify a driver’s head pose from naturalistic videos, which were captured by a camera providing a side view, instead of directly facing the driver. These approaches employ a sequence of five key points on the driver’s face. We compare these three proposed approaches with each other as well as with three different baselines by using leave-one-driver-out cross-validation on nine different drivers. Results show that our proposed method employing a Bidirectional Gated Recurrent Unit (BiGRU) outperforms the best performing baseline by 11% in terms of overall accuracy.

Distracted or drowsy driving is unsafe driving behavior responsible for thousands of crashes every year. Studying driver behavior has challenges associated with observing drivers in their natural environment. The naturalistic driving study (NDS) has become the most sought-after approach, since it eliminates the bias of a controlled setup, allowing researchers to understand drivers’ behavior in real-world scenarios. Video recordings collected in NDS research are incredibly insightful in identifying driver errors. Computer vision techniques have been used to autonomously analyze video data and classify drivers’ behavior. While computer vision scientists focus on image analytics, NDS researchers are interested in the factors impacting driver behavior. This survey paper makes a concerted effort to serve both communities by comprehensively reviewing studies, describing their data collection, computer vision techniques implemented, and performance in classifying driver behavior. The scope is limited to studies employing at least one camera observing the driver inside a vehicle. Based on their objective, papers have been classified as detecting low-level (e.g. head orientation) or high-level (e.g. distraction detection) driver information. Papers have been further classified based on the datasets they employ. In addition to twelve public datasets, many private datasets have also been identified, and their data collection design is discussed to highlight any impact on model performance. Across each task, algorithms employed and their performance are discussed to establish a baseline. A comparison of different frameworks for NDS video data analytics throws light on the existing gaps in the state-of-the-art that can be addressed by future computer vision research.

Our goal is to improve driver safety predictions in at-risk medical or aging populations from naturalistic driving video data. To meet this goal, we developed a novel model capable of detecting and tracking unsafe lane departure events (e.g., changes and incursions), which may occur more frequently in at-risk driver populations. The model detects and tracks roadway lane markings in challenging, low-resolution driving videos using a semantic lane detection pre-processor (Mask R-CNN) utilizing the driver's forward lane region, demarking the convex hull that represents the driver's lane. The hull centroid is tracked over time, improving lane tracking over approaches which detect lane markers from single video frames. The lane time series was denoised using a Fix-lag Kalman filter. Preliminary results show promise for robust lane departure event detection. Overall recall for detecting lane departure events was 81.82%. The F1 score was 75% (precision 69.23%) and 70.59% (precision 62.07%) for left and right lane departures, respectively. Future investigations include exploring (1) horizontal offset as a means to detect lead vehicle proximity, even when image perspectives are known to have a chirp effect and (2) Long Short Term Memory (LSTM) models to detect peaks instead of a peak detection algorithm.

Reliable estimation of performance measures such as queue and delay at intersections is important for the proper management of traffic. The information about these variables is valuable for the development of various traffic control strategies. The spatial nature of queue and delay makes their direct measurement a challenging task. The present study estimated these performance measures for the scenario when the queue ends within the advance detector using the data obtained from loop detectors installed at the entry and the exit of the intersection. A detailed analysis of the data obtained from loop detectors revealed that there were errors in the data. Two model based schemes, namely the occupancy based method and the queue clearance based method, were used for estimation of queue and delay using the erroneous data obtained from loop detectors. The results showed that the queue clearance based method was performing better while estimating queue and delay compared to the occupancy based method. Thus, the queue clearance based method would be valuable for the estimation of queues and delays while implementing with erroneous field data.