Abstract

Effective decision-making in complex environments with multi-discrete action spaces poses significant challenges for agent architectures, particularly in image-based settings. While Decision Transformers have shown promise in various domains, their performance often suffers in environments where agents must handle multi-discrete actions. Existing enhancements to Decision Transformer architectures have yet to address this critical issue, limiting their ability to support agents in learning robust policies in these environments. To address this gap, we propose Multi-State Action Tokenisation (M-SAT), a novel approach designed to improve agent decision-making by tokenising actions at the individual action level and incorporating auxiliary state information. This disentanglement of actions improves both the performance of agents and the interpretability of individual actions within attention layers, fostering better visibility into agent decision processes. Importantly, M-SAT facilitates the development of more interpretable and transparent agents capable of making complex decisions in dynamic environments involving multi-discrete action spaces. We evaluate M-SAT on the challenging ViZDoom environments, focusing on scenarios with multi-discrete action spaces and image-based observations, such as Deadly Corridor, My Way Home and Death Match. Our approach demonstrates superior performance compared to baseline Decision Transformers, with no additional data or significant computational overheads. Furthermore, we observe that M-SAT does not require positional encoding to achieve high performance, with its removal occasionally leading to further improvements. These findings suggest that M-SAT enables more efficient and interpretable agent-based decision-making in multi-discrete action spaces.

Abstract

The study of information revivals, witnessing the violation of certain data-processing inequalities, has provided an important paradigm in the study of non-Markovian quantum stochastic processes. Although often used interchangeably, we argue here that the notions of “revivals” and “backflows,” i.e., flows of information from the environment back into the system, are distinct: an information revival can occur without any backflow ever taking place. In this paper, we examine in detail the phenomenon of noncausal revivals and relate them to the theory of short Markov chains and squashed non-Markovianity. We also provide an operational condition, in terms of system-only degrees of freedom, to witness the presence of genuine backflow that cannot be explained by noncausal revivals. As a byproduct, we demonstrate that focusing on processes with genuine backflows, while excluding those with only noncausal revivals, resolves the issue of nonconvexity of Markovianity, thus enabling the construction of a convex resource theory of genuine quantum non-Markovianity

Abstract

An influential mathematical model of memory, the temporal context model (TCM), posits that we encode items and their associations with temporal context (Howard & Kahana, 2002). Temporal context is conceived of as a recency-weighted average of past experiences. Critically, the model assumes that when an item is retrieved later, the associated temporal context is also obligatorily retrieved. Existing evidence for the idea of retrieved temporal context primarily comes from free-recall studies. However, free recall introduces some critical confounds that are difficult to resolve (Folkerts et al., 2018) and also encourages memory strategies that may mimic temporal context effects (Hintzman, 2011). Schwartz et al. (2005) showed that temporal context influences image recognition within short experimental lists. We extend this by using the Natural Scenes Dataset (NSD) to demonstrate that reinstating temporal context enhances recognition accuracy even across long timescales. We demonstrate that when the relevant temporal context is successfully reinstated for a reference image, the recognition accuracy of subsequent images increases. Critically, we show that this influence falls off with temporal distance at encoding from the reference image only when the temporal context is successfully retrieved, as predicted by TCM. Furthermore, the slope of this temporal gradient increases as a function of the strength of the influence of the retrieved temporal context. These findings extend our understanding of temporal context effects in episodic memory by showing that temporal context is retrieved even in tasks that do not encourage linking between items as a memory strategy.

Abstract

A demonstration of the implementation of vehicle occupancy detection on hardware-software is shown in this paper. For the purpose of validating applications for vehicle occupancy detection, a hardware Field Programmable Gate Array (FPGA) platform, also known as PYNQ-ZU, is a feasible embedded architecture. Automatic in-car occupancy monitoring is an important technology in modern transportation, with major implications for safety, energy efficiency, and smart vehicle management. One of the primary benefits of mmWave radar is its ability to accurately detect the number and location of vehicle occupants, mmWave radar ensures robust detection under all lighting and weather conditions. In our research, the proposed approach was applied to point cloud images. Following the generation of 3D point cloud images, two filters, Top-View (TV), and Front-View (FV), were used to improve vehicle occupancy detection. These filters transformed 3D images into 2D ones. TV filter was found to be more effective than the FV filter. After filtering the 2D images, Mexican Hat Wavelet Transform (MHWT) was used to extract features from them. Four machine learning methods were then used to determine vehicle seat occupancy, with Logistic Regression (LR) and Support Vector Machine (SVM) producing the highest results, with an accuracy of 98%. In comparison to existing methods, the proposed approach, which utilizes mmWave radar, TV Filter, MHWT, FPGA (PYNQ-ZU), and LR, was determined to significantly improve the accuracy of vehicle occupancy detection.

Abstract

A graph transactional database (GTD) is a collection of graphs. Frequent Top-k Subgraph Pattern Mining (FTSPM) involves finding the complete set of top-k frequently occurring subgraph patterns in a GTD. Most previous studies focused on finding these patterns in certain graphs by disregarding the crucial information regarding the existential probabilities that may exist between the edges of any two nodes. With this motivation, this paper proposes a novel model to discover top-k (frequently occurring) subgraphs in an uncertain graph transactional database. We introduce a novel algorithm, top-k uncertain subgraph miner (TUSM), to find all top-k subgraphs in the data. We also introduce an approximate top-k uncertain subgraph miner (ATUSM) algorithm to tackle the computational expansiveness of TUSM. Experimental results on synthetic and protein-protein interaction datasets demonstrate that the proposed model finds valuable information and the algorithms are efficient.

Abstract

We develop new algorithms for Quantum Singular Value Transformation (QSVT), a unifying framework underlying a wide range of quantum algorithms. Existing implementations of QSVT rely on block encoding, incurring $O(log L)$ ancilla overhead and circuit depth $widetilde{O}(d lambda L)$ for polynomial transformations of a Hamiltonian $H = sum_{k=1}^L lambda_k H_k$, where $d$ is the polynomial degree, and $lambda = sum_k |lambda_k|$. We introduce a new approach that eliminates block encoding, needs only a single ancilla qubit, and maintains near-optimal complexity, using only basic Hamiltonian simulation methods such as Trotterization. Our method achieves a circuit depth of $widetilde{O}left(L (d lambda_{text{comm}})^{1+o(1)}right)$, without any multi-qubit controlled gates. Here, $lambda_{text{comm}}$ depends on the nested commutators of the $H_k$'s and can be much smaller than $lambda$. Central to our technique is a novel use of Richardson extrapolation, enabling systematic error cancellation in interleaved sequences of arbitrary unitaries and Hamiltonian evolution operators, establishing a broadly applicable framework beyond QSVT. Additionally, we propose two randomized QSVT algorithms for cases with only sampling access to Hamiltonian terms. The first uses qDRIFT, while the second replaces block encodings in QSVT with randomly sampled unitaries. Both achieve quadratic complexity in $d$, which we establish as a lower bound for any randomized method implementing polynomial transformations in this model. Finally, as applications, we develop end-to-end quantum algorithms for quantum linear systems and ground state property estimation, achieving near-optimal complexity without oracular access. Our results provide a new framework for quantum algorithms, reducing hardware overhead while maintaining near-optimal performance, with implications for both near-term and fault-tolerant quantum computing.

Abstract

Recent evaluations of LLMs on coreference resolution have revealed that traditional output formats and evaluation metrics do not fully capture the models' referential understanding. To address this, we introduce IdentifyMe, a new benchmark for mention resolution presented in a multiple-choice question (MCQ) format, commonly used for evaluating LLMs. IdentifyMe features long narratives and employs heuristics to exclude easily identifiable mentions, creating a more challenging task. The benchmark also consists of a curated mixture of different mention types and corresponding entities, allowing for a fine-grained analysis of model performance. We evaluate both closed- and open source LLMs on IdentifyMe and observe a significant performance gap (20-30%) between the state-of-the-art sub-10B open models vs. closed ones. We observe that pronominal mentions, which have limited surface information, are typically much harder for models to resolve than nominal mentions. Additionally, we find that LLMs often confuse entities when their mentions overlap in nested structures. The highest-scoring model, GPT-4o, achieves 81.9% accuracy, highlighting the strong referential capabilities of state-of-the-art LLMs while also indicating room for further improvement

Abstract

Parkinson’s Disease (PD) neurodegenerative disorder which lacks reliable early diagnostic tests. In this study, we present a method for PD prediction that uses multimodal data from the Parkinson Progression Marker Initiative (PPMI), specifically clinico-demographic, biospecimen, and genetic data, to improve predictive accuracy and facilitate timely interventions via a multimodal machine learning approach. We evaluated data obtained from 598 participants (171 healthy controls and 427 PD patients) in three different modalities: 29 clinical features, five cerebrospinal fluid biomarkers, and 154 SNPs (single nucleotide polymorphisms), which were selected through a biology-driven feature selection method. We utilized three multimodal integration strategies—early, intermediate, and late—and trained various machine learning models, including LightGBM and Multilayer Perceptron (MLP), each of which was optimized by hyperparameter tuning and cross-validation. In early integration, we combined feature sets from all modalities into a single set, leveraging complementary information to increase predictive power. The intermediate integration method made use of autoencoders to encode features into a single 12 dimensional vector as the input into a Neural Network classifier. Late integration combined outputs from the top-performing models for each modality using ensemble techniques such as Voting Classifier and Stacking. We found that early integration achieved the highest performance, with Support Vector Machines with 90% accuracy, 0.98 AUC-ROC, 0.99 precision, and 0.93 F1 score. Intermediate integration with the Neural Network classifier followed with an AUC-ROC of 0.84 and an F1-score of 0.85. Our feature importance analysis identified two clinical scores, namely UPSIT (related to sensory decline) and SCOPA (measuring autonomic dysfunction), along with two SNPs (chr4 90755939 A G and chr4 90646886 G A, both previously linked to PD susceptibility) as crucial predictors, emphasizing their well established relevance in PD diagnosis. Early diagnosis and treatment of PD patients are facilitated by the integration of multiple data modalities, which also greatly increases the predicted accuracy for the disease while also providing us with a thorough understanding of its complexity.

Abstract

The constant shifts in social and political contexts, driven by emerging social movements and political events, lead to new forms of hate content and previously unrecognized hate patterns that machine learning models may not have captured. Some recent literature proposes data augmentation-based techniques to enrich existing hate datasets by incorporating samples that reveal new implicit hate patterns. This approach aims to improve the model's performance on out-of-domain implicit hate instances. It is observed, that further addition of more samples for augmentation results in the decrease of the performance of the model. In this work, we propose a Knowledge Transfer-driven Concept Refinement method that distills and refines the concepts related to implicit hate samples through novel prototype alignment and concept losses, alongside data augmentation based on concept activation vectors. Experiments with several publicly available datasets show that incorporating additional implicit samples reflecting new hate patterns through concept refinement enhances the model's performance, surpassing baseline results while maintaining cross-dataset generalization capabilities.

Abstract

In this work, we present IDD-CRS, a large-scale dataset focused on critical road scenarios, captured using Advanced Driver Assistance Systems (ADAS) and dash cameras. Unlike existing datasets that predominantly emphasize pedestrian safety and vehicle safety separately, IDD-CRS incorporates both vehicle and pedestrian behaviors, offering a more comprehensive view of road safety. The dataset includes diverse scenarios, such as high-speed lane changes, unsafe vehicle approaches to pedestrians and cyclists, and complex interactions between ego vehicles and other road agents. Leveraging ADAS technology allows us to accurately define the temporal boundaries of actions, resulting in precise annotations and more reliable safety analysis. With 90 hours of video footage, consisting of 5400 one-minute-long videos and 135,000 frames, IDD-CRS introduces new vehicle related classes and hard negative classes, establishing baselines for action recognition and long-tail action recognition tasks. Our benchmarks reveal the limitations of current models, pointing toward future advancements needed for improving road safety technology.