Policy gradient methods have enabled deep reinforcement learning (RL) to approach challenging continuous control problems, even when the underlying systems involve highly nonlinear dynamics that generate complex non-smooth optimization landscapes. We develop a rigorous framework for understanding how policy gradient methods mollify non-smooth optimization landscapes to enable effective policy search, as well as the downside of it: while making the objective function smoother and easier to optimize, the stochastic objective deviates further from the original problem. We demonstrate the equivalence between policy gradient methods and solving backward heat equations. Following the ill-posedness of backward heat equations from PDE theory, we present a fundamental challenge to the use of policy gradient under stochasticity. Moreover, we make the connection between this limitation and the uncertainty principle in harmonic analysis to understand the effects of exploration with stochastic policies in RL. We also provide experimental results to illustrate both the positive and negative aspects of mollification effects in practice.

Key-value~(KV) caching is an important technique to accelerate the inference of large language models~(LLMs), but incurs significant memory overhead. To compress the size of KV cache, existing methods often compromise precision or require extra data for calibration, limiting their practicality in LLM deployment. In this paper, we introduce \textbf{DecoQuant}, a novel data-free low-bit quantization technique based on tensor decomposition methods, to effectively compress KV cache. Our core idea is to adjust the outlier distribution of the original matrix by performing tensor decomposition, so that the quantization difficulties are migrated from the matrix to decomposed local tensors. Specially, we find that outliers mainly concentrate on small local tensors, while large tensors tend to have a narrower value range. Based on this finding, we propose to apply low-bit quantization to the large tensor, while maintaining high-precision representation for the small tensor. Furthermore, we utilize the proposed quantization method to compress the KV cache of LLMs to accelerate the inference and develop an efficient dequantization kernel tailored specifically for DecoQuant. Through extensive experiments, DecoQuant demonstrates remarkable efficiency gains, showcasing up to a $\sim$75\% reduction in memory footprint while maintaining comparable generation quality.

Deploying large language models (LLMs) poses challenges in terms of resource limitations and inference efficiency. To address these challenges, recent research has focused on using smaller task-specific language models, which are enhanced by distilling the knowledge rationales generated by LLMs. However, previous works mostly emphasize the effectiveness of positive knowledge, while overlooking the knowledge noise and the exploration of negative knowledge. In this paper, we first propose a general approach called quality-guided contrastive rationale distillation for reasoning capacity learning, considering contrastive learning perspectives. For the learning of positive knowledge, we collect positive rationales through self-consistency to denoise the LLM rationales generated by temperature sampling. For the negative knowledge distillation, we generate negative rationales using temperature sampling for the iteration-before smaller language models themselves. Finally, a contrastive loss is designed to better distill the positive and negative rationales into the smaller language model, where an online-update discriminator is used to judge the qualities of rationales and assign weights for better optimizing the training process. Through extensive experiments on multiple reasoning tasks, we demonstrate that our method consistently outperforms the previous distillation methods and produces higher-quality rationales.

Federated learning achieves effective performance in modeling decentralized data. In practice, client data are not well-labeled, which makes it potential for federated unsupervised learning (FUSL) with non-IID data. However, the performance of existing FUSL methods suffers from insufficient representations, i.e., (1) representation collapse entanglement among local and global models, and (2) inconsistent representation spaces among local models. The former indicates that representation collapse in local model will subsequently impact the global model and other local models. The latter means that clients model data representation with inconsistent parameters due to the deficiency of supervision signals. In this work, we propose FedU2 which enhances generating uniform and unified representation in FUSL with non-IID data. Specifically, FedU2 consists of flexible uniform regularizer (FUR) and efficient unified aggregator (EUA). FUR in each client avoids representation collapse via dispersing samples uniformly, and EUA in server promotes unified representation by constraining consistent client model updating. To extensively validate the performance of FedU2, we conduct both cross-device and cross-silo evaluation experiments on two benchmark datasets, i.e., CIFAR10 and CIFAR100.

We propose a computational graph representation of high-order Feynman diagrams in Quantum Field Theory (QFT), applicable to any combination of spatial, temporal, momentum, and frequency domains. Utilizing the Dyson-Schwinger and parquet equations, our approach effectively organizes these diagrams into a fractal structure of tensor operations, significantly reducing computational redundancy. This approach not only streamlines the evaluation of complex diagrams but also facilitates an efficient implementation of the field-theoretic renormalization scheme, crucial for enhancing perturbative QFT calculations. Key to this advancement is the integration of Taylor-mode automatic differentiation, a key technique employed in machine learning packages to compute higher-order derivatives efficiently on computational graphs. To operationalize these concepts, we develop a Feynman diagram compiler that optimizes diagrams for various computational platforms, utilizing machine learning frameworks. Demonstrating this methodology's effectiveness, we apply it to the three-dimensional uniform electron gas problem, achieving unprecedented accuracy in calculating the quasiparticle effective mass at metal density. Our work demonstrates the synergy between QFT and machine learning, establishing a new avenue for applying AI techniques to complex quantum many-body problems.

We introduce a privacy auditing scheme for ML models that relies on membership inference attacks using generated data as "non-members". This scheme, which we call PANORAMIA, quantifies the privacy leakage for large-scale ML models without control of the training process or model re-training and only requires access to a subset of the training data. To demonstrate its applicability, we evaluate our auditing scheme across multiple ML domains, ranging from image and tabular data classification to large-scale language models.

We present a sensor misalignment-tolerant AUV navigation method that leverages measurements from an acoustic array and dead reckoned information. Recent studies have demonstrated the potential use of passive acoustic Direction of Arrival (DoA) measurements for AUV navigation without requiring ranging measurements. However, the sensor misalignment between the acoustic array and the attitude sensor was not accounted for. Such misalignment may deteriorate the navigation accuracy. This paper proposes a novel approach that allows simultaneous AUV navigation, beacon localization, and sensor alignment. An Unscented Kalman Filter (UKF) that enables the necessary calculations to be completed at an affordable computational load is developed. A Nonlinear Least Squares (NLS)-based technique is employed to find an initial solution for beacon localization and sensor alignment as early as possible using a short-term window of measurements. Experimental results demonstrate the performance of the proposed method.

Multi-modal large language models have demonstrated impressive performance across various tasks in different modalities. However, existing multi-modal models primarily emphasize capturing global information within each modality while neglecting the importance of perceiving local information across modalities. Consequently, these models lack the ability to effectively understand the fine-grained details of input data, limiting their performance in tasks that require a more nuanced understanding. To address this limitation, there is a compelling need to develop models that enable fine-grained understanding across multiple modalities, thereby enhancing their applicability to a wide range of tasks. In this paper, we propose GroundingGPT, a language enhanced multi-modal grounding model. Beyond capturing global information like other multi-modal models, our proposed model excels at tasks demanding a detailed understanding of local information within the input. It demonstrates precise identification and localization of specific regions in images or moments in videos. To achieve this objective, we design a diversified dataset construction pipeline, resulting in a multi-modal, multi-granularity dataset for model training. The code, dataset, and demo of our model can be found at https: //github.com/lzw-lzw/GroundingGPT.

Saliency maps have proven to be a highly efficacious approach for explicating the decisions of Convolutional Neural Networks. However, extant methodologies predominantly rely on gradients, which constrain their ability to explicate complex models. Furthermore, such approaches are not fully adept at leveraging negative gradient information to improve interpretive veracity. In this study, we present a novel concept, termed positive and negative excitation, which enables the direct extraction of positive and negative excitation for each layer, thus enabling complete layer-by-layer information utilization sans gradients. To organize these excitations into final saliency maps, we introduce a double-chain backpropagation procedure. A comprehensive experimental evaluation, encompassing both binary classification and multi-classification tasks, was conducted to gauge the effectiveness of our proposed method. Encouragingly, the results evince that our approach offers a significant improvement over the state-of-the-art methods in terms of salient pixel removal, minor pixel removal, and inconspicuous adversarial perturbation generation guidance. Additionally, we verify the correlation between positive and negative excitations.

This study presents a novel approach to addressing the challenge of missing data in multivariate time series, with a particular focus on the complexities of healthcare data. Our Conditional Self-Attention Imputation (CSAI) model, grounded in a transformer-based framework, introduces a conditional hidden state initialization tailored to the intricacies of medical time series data. This methodology diverges from traditional imputation techniques by specifically targeting the imbalance in missing data distribution, a crucial aspect often overlooked in healthcare datasets. By integrating advanced knowledge embedding and a non-uniform masking strategy, CSAI adeptly adjusts to the distinct patterns of missing data in Electronic Health Records (EHRs).