Search

Meta-reinforcement learning (Meta-RL) facilitates rapid adaptation to unseen tasks but faces challenges in long-horizon environments. Skill-based approaches tackle this by decomposing state-action sequences into reusable skills and employing hierarchical decision-making. However, these methods are highly susceptible to noisy offline demonstrations, resulting in unstable skill learning and degraded performance. To overcome this, we propose Prioritized Refinement for Skill-Based Meta-RL (PRISM), a robust framework that integrates exploration near noisy data to generate online trajectories and combines them with offline data. Through prioritization, PRISM extracts high-quality data to learn task-relevant skills effectively. By addressing the impact of noise, our method ensures stable skill learning and achieves superior performance in long-horizon tasks, even with noisy and sub-optimal data., Comment: 8 pages main, 19 pages appendix with reference. Submitted to ICML 2025

Recent large language models (LLMs) with enormous model sizes use many GPUs to meet memory capacity requirements incurring substantial costs for token generation. To provide cost-effective LLM inference with relaxed latency constraints, extensive research has focused on expanding GPU memory by leveraging the host memory. However, LLM inference engines that utilize the host memory often face underutilization of GPU compute units, as a considerable portion of inference time is spent in loading the model onto the GPU via host-GPU interconnect. To tackle these challenges of the host memory offloading for LLM, we introduce HybridServe, an LLM inference system with activation checkpointing based on activation caching. The activation cache stores activation checkpoints generated during intermediate inference stages, allowing the fast recomputation of KV cache while model parameters are transferred to GPU from host memory. Unlike conventional methods that recompute the KV cache from scratch using token IDs, the activation cache allows bypassing projection and FFN operations. To balance between the activation recomputation and parameter loading overhead, this study proposes a KV-activation hybrid caching scheme which finds the best ratio of the key-value and activation caches to adjust the recomputation time. Our system achieves 2.19x throughput improvement over the state-of-the-art prior work for offloading both model weights and KV cache., Comment: 14 pages, 15 figures

Point-based interactive colorization techniques allow users to effortlessly colorize grayscale images using user-provided color hints. However, point-based methods often face challenges when different colors are given to semantically similar areas, leading to color intermingling and unsatisfactory results-an issue we refer to as color collapse. The fundamental cause of color collapse is the inadequacy of points for defining the boundaries for each color. To mitigate color collapse, we introduce a lasso tool that can control the scope of each color hint. Additionally, we design a framework that leverages the user-provided lassos to localize the attention masks. The experimental results show that using a single lasso is as effective as applying 4.18 individual color hints and can achieve the desired outcomes in 30% less time than using points alone., Comment: Accepted to AAAI2025

Recent advancements in computer vision have led to a renewed interest in developing assistive technologies for individuals with visual impairments. Although extensive research has been conducted in the field of computer vision-based assistive technologies, most of the focus has been on understanding contexts in images, rather than addressing their physical safety and security concerns. To address this challenge, we propose the first step towards detecting anomalous situations for visually impaired people by observing their entire surroundings using an egocentric 360-degree camera. We first introduce a novel egocentric 360-degree video dataset called VIEW360 (Visually Impaired Equipped with Wearable 360-degree camera), which contains abnormal activities that visually impaired individuals may encounter, such as shoulder surfing and pickpocketing. Furthermore, we propose a new architecture called the FDPN (Frame and Direction Prediction Network), which facilitates frame-level prediction of abnormal events and identifying of their directions. Finally, we evaluate our approach on our VIEW360 dataset and the publicly available UCF-Crime and Shanghaitech datasets, demonstrating state-of-the-art performance., Comment: WACV2025

Constraint-based offline reinforcement learning (RL) involves policy constraints or imposing penalties on the value function to mitigate overestimation errors caused by distributional shift. This paper focuses on a limitation in existing offline RL methods with penalized value function, indicating the potential for underestimation bias due to unnecessary bias introduced in the value function. To address this concern, we propose Exclusively Penalized Q-learning (EPQ), which reduces estimation bias in the value function by selectively penalizing states that are prone to inducing estimation errors. Numerical results show that our method significantly reduces underestimation bias and improves performance in various offline control tasks compared to other offline RL methods, Comment: 10 technical page followed by references and appendix. Accepted to Neurips 2024 as spotlight paper

Modern transformer-based Large Language Models (LLMs) are constructed with a series of decoder blocks. Each block comprises three key components: (1) QKV generation, (2) multi-head attention, and (3) feed-forward networks. In batched processing, QKV generation and feed-forward networks involve compute-intensive matrix-matrix multiplications (GEMM), while multi-head attention requires bandwidth-heavy matrix-vector multiplications (GEMV). Machine learning accelerators like TPUs or NPUs are proficient in handling GEMM but are less efficient for GEMV computations. Conversely, Processing-in-Memory (PIM) technology is tailored for efficient GEMV computation, while it lacks the computational power to handle GEMM effectively. Inspired by this insight, we propose NeuPIMs, a heterogeneous acceleration system that jointly exploits a conventional GEMM-focused NPU and GEMV-optimized PIM devices. The main challenge in efficiently integrating NPU and PIM lies in enabling concurrent operations on both platforms, each addressing a specific kernel type. First, existing PIMs typically operate in a "blocked" mode, allowing only either NPU or PIM to be active at any given time. Second, the inherent dependencies between GEMM and GEMV in LLMs restrict their parallel processing. To tackle these challenges, NeuPIMs is equipped with dual row buffers in each bank, facilitating the simultaneous management of memory read/write operations and PIM commands. Further, NeuPIMs employs a runtime sub-batch interleaving technique to maximize concurrent execution, leveraging batch parallelism to allow two independent sub-batches to be pipelined within a single NeuPIMs device. Our evaluation demonstrates that compared to GPU-only, NPU-only, and a na\"ive NPU+PIM integrated acceleration approaches, NeuPIMs achieves 3$\times$, 2.4$\times$ and 1.6$\times$ throughput improvement, respectively., Comment: 16 pages, 15 figures

The smooth quintic del Pezzo variety $Y$ is well-known to be obtained as a linear sections of the Grassmannian variety $\mathrm{Gr}(2,5)$ under the Pl\"ucker embedding into $\mathbb{P}^{9}$. Through a local computation, we show the Hilbert scheme of conics in $Y$ for $\text{dim} Y \ge 3$ can be obtained from a certain Grassmannian bundle by a single blowing up/down transformation.

Point-interactive image colorization aims to colorize grayscale images when a user provides the colors for specific locations. It is essential for point-interactive colorization methods to appropriately propagate user-provided colors (i.e., user hints) in the entire image to obtain a reasonably colorized image with minimal user effort. However, existing approaches often produce partially colorized results due to the inefficient design of stacking convolutional layers to propagate hints to distant relevant regions. To address this problem, we present iColoriT, a novel point-interactive colorization Vision Transformer capable of propagating user hints to relevant regions, leveraging the global receptive field of Transformers. The self-attention mechanism of Transformers enables iColoriT to selectively colorize relevant regions with only a few local hints. Our approach colorizes images in real-time by utilizing pixel shuffling, an efficient upsampling technique that replaces the decoder architecture. Also, in order to mitigate the artifacts caused by pixel shuffling with large upsampling ratios, we present the local stabilizing layer. Extensive quantitative and qualitative results demonstrate that our approach highly outperforms existing methods for point-interactive colorization, producing accurately colorized images with a user's minimal effort. Official codes are available at https://pmh9960.github.io/research/iColoriT, Comment: Accepted to WACV 2023

By the reduced component in a moduli space of stable quasimaps to n-dimensional projective space we mean the closure of the locus in which the domain curves are smooth. As in the moduli space of stable maps, we prove the reduced component is smooth in genus 2, degree greater or equal to 3. Then we prove the virtual fundamental cycle of the moduli space of stable quasimaps to a complete intersection X in the projective space of genus 2, degree greater or equal to 3 is explicitly expressed in terms of the fundamental cycle of the reduced component of the projective space and virtual cycles of lower genus moduli spaces of X., Comment: 21pages

We compute Gromov-Witten (GW) and Donaldson-Thomas (DT) invariants (and also descendant invariants) for local CY 4-folds over Fano 3-folds, V_5 and V_22 up to degree 3. We use torus localization for GW invariants computation, and use classical results for Hilbert schemes on V_5 and V_22 for DT invariants computation. From these computations, one can check correspondence between DT and Gopakumar-Vafa (GV) invariants conjectured by Cao-Maulik-Toda in genus 0. Also we can compute genus 1 GV invariants via the conjecture of Cao-Toda, which turned out to be 0. These fit into the fact that there are no smooth elliptic curves in V_5 and V_22 up to degree 3., Comment: 20pages, all comments are welcome!

The widening availability of hardware-based trusted execution environments (TEEs) has been accelerating the adaptation of new applications using TEEs. Recent studies showed that a cloud application consists of multiple distributed software modules provided by mutually distrustful parties. The applications use multiple TEEs (enclaves) communicating through software-encrypted memory channels. Such execution model requires bi-directional protection: protecting the rest of the system from the enclave module with sandboxing and protecting the enclave module from a third-part module and operating systems. However, the current TEE model, such as Intel SGX, cannot efficiently represent such distributed sandbox applications. To overcome the lack of hardware supports for sandboxed TEEs, this paper proposes an extended enclave model called Stockade, which supports distributed sandboxes hardened by hardware. Stockade proposes new three key techniques. First, it extends the hardware-based memory isolation in SGX to confine a user software module only within its enclave. Second, it proposes a trusted monitor enclave that filters and validates systems calls from enclaves. Finally, it allows hardware-protected memory sharing between a pair of enclaves for efficient protected communication without software-based encryption. Using an emulated SGX platform with the proposed extensions, this paper shows that distributed sandbox applications can be effectively supported with small changes of SGX hardware.

Memory disaggregation provides efficient memory utilization across network-connected systems. It allows a node to use part of memory in remote nodes in the same cluster. Recent studies have improved RDMA-based memory disaggregation systems, supporting lower latency and higher bandwidth than the prior generation of disaggregated memory. However, the current disaggregated memory systems manage remote memory only at coarse granularity due to the limitation of the access validation mechanism of RDMA. In such systems, to support fine-grained remote page allocation, the trustworthiness of all participating systems needs to be assumed, and thus a security breach in a node can propagate to the entire cluster. From the security perspective, the memory-providing node must protect its memory from memory-requesting nodes. On the other hand, the memory-requesting node requires the confidentiality and integrity protection of its memory contents even if they are stored in remote nodes. To address the weak isolation support in the current system, this study proposes a novel hardware-assisted memory disaggregation system. Based on the security features of FPGA, the logic in each per-node FPGA board provides a secure memory disaggregation engine. With its own networks, a set of FPGA-based engines form a trusted memory disaggregation system, which is isolated from the privileged software of each participating node. The secure memory disaggregation system allows fine-grained memory management in memory-providing nodes, while the access validation is guaranteed with the hardware-hardened mechanism. In addition, the proposed system hides the memory access patterns observable from remote nodes, supporting obliviousness. Our evaluation with FPGA implementation shows that such fine-grained secure disaggregated memory is feasible with comparable performance to the latest software-based techniques.

Deep neural networks for automatic image colorization often suffer from the color-bleeding artifact, a problematic color spreading near the boundaries between adjacent objects. Such color-bleeding artifacts debase the reality of generated outputs, limiting the applicability of colorization models in practice. Although previous approaches have attempted to address this problem in an automatic manner, they tend to work only in limited cases where a high contrast of gray-scale values are given in an input image. Alternatively, leveraging user interactions would be a promising approach for solving this color-breeding artifacts. In this paper, we propose a novel edge-enhancing network for the regions of interest via simple user scribbles indicating where to enhance. In addition, our method requires a minimal amount of effort from users for their satisfactory enhancement. Experimental results demonstrate that our interactive edge-enhancing approach effectively improves the color-bleeding artifacts compared to the existing baselines across various datasets., Comment: Accepted to ICCV 2021 as oral presentation (3% acceptance rate)

The objective of this study was to investigate the importance of multiple county-level features in the trajectory of COVID-19. We examined feature importance across 2,787 counties in the United States using a data-driven machine learning model. We trained random forest models using 23 features representing six key influencing factors affecting pandemic spread: social demographics of counties, population activities, mobility within the counties, movement across counties, disease attributes, and social network structure. Also, we categorized counties into multiple groups according to their population densities, and we divided the trajectory of COVID-19 into three stages: the outbreak stage, the social distancing stage, and the reopening stage. The study aims to answer two research questions: (1) The extent to which the importance of heterogeneous features evolves in different stages; (2) The extent to which the importance of heterogeneous features varies across counties with different characteristics. We fitted a set of random forest models to determine weekly feature importance. The results showed that: (1) Social demographic features, such as gross domestic product, population density, and minority status maintained high-importance features throughout stages of COVID-19 across the 2787 studied counties; (2) Within-county mobility features had the highest importance in county clusters with higher population densities; (3) The feature reflecting the social network structure (Facebook, social connectedness index), had higher importance in the models for counties with higher population densities. The results show that the data-driven machine learning models could provide important insights to inform policymakers regarding feature importance for counties with various population densities and in different stages of a pandemic life cycle.

The spread of pandemics such as COVID-19 is strongly linked to human activities. The objective of this paper is to specify and examine early indicators of disease spread risk in cities during the initial stages of outbreak based on patterns of human activities obtained from digital trace data. In this study, the Venables distance (D_v), and the activity density (D_a) are used to quantify and evaluate human activities for 193 US counties, whose cumulative number of confirmed cases was greater than 100 as of March 31, 2020. Venables distance provides a measure of the agglomeration of the level of human activities based on the average distance of human activities across a city or a county (less distance could lead to a greater contact risk). Activity density provides a measure of level of overall activity level in a county or a city (more activity could lead to a greater risk). Accordingly, Pearson correlation analysis is used to examine the relationship between the two human activity indicators and the basic reproduction number in the following weeks. The results show statistically significant correlations between the indicators of human activities and the basic reproduction number in all counties, as well as a significant leader-follower relationship (time lag) between them. The results also show one to two weeks' lag between the change in activity indicators and the decrease in the basic reproduction number. This result implies that the human activity indicators provide effective early indicators for the spread risk of the pandemic during the early stages of the outbreak. Hence, the results could be used by the authorities to proactively assess the risk of disease spread by monitoring the daily Venables distance and activity density in a proactive manner., Comment: 12 pages, 8 figures

The rapid spread of COVID-19 in the United States has imposed a major threat to public health, the real economy, and human well-being. With the absence of effective vaccines, the preventive actions of social distancing and travel reduction are recognized as essential non-pharmacologic approaches to control the spread of COVID-19. Prior studies demonstrated that human movement and mobility drove the spatiotemporal distribution of COVID-19 in China. Little is known, however, about the patterns and effects of co-location reduction on cross-county transmission risk of COVID-19. This study utilizes Facebook co-location data for all counties in the United States from March to early May 2020. The analysis examines the synchronicity and time lag between travel reduction and pandemic growth trajectory to evaluate the efficacy of social distancing in ceasing the population co-location probabilities, and subsequently the growth in weekly new cases. The results show that the mitigation effects of co-location reduction appear in the growth of weekly new cases with one week of delay. Furthermore, significant segregation is found among different county groups which are categorized based on numbers of cases. The results suggest that within-group co-location probabilities remain stable, and social distancing policies primarily resulted in reduced cross-group co-location probabilities (due to travel reduction from counties with large number of cases to counties with low numbers of cases). These findings could have important practical implications for local governments to inform their intervention measures for monitoring and reducing the spread of COVID-19, as well as for adoption in future pandemics. Public policy, economic forecasting, and epidemic modeling need to account for population co-location patterns in evaluating transmission risk of COVID-19 across counties., Comment: 12 pages, 7 figures

In our previous work, we provided an algebraic proof of the Zinger's comparison formula between genus one Gromov-Witten invariants and reduced invariants when the target space is a complete intersection of dimension two or three in a projective space. In this paper, we extend the result in any dimensions and for descendant invariants., Comment: 29 pages, all comments are welcome!

Let $C$ be an algebraic curve of genus $g$ and $L$ a line bundle over $C$. Let $\mathcal{MS}_C(n,L)$ and $\mathcal{MO}_C(n,L)$ be the moduli spaces of $L$-valued symplectic and orthogonal bundles respectively, over $C$ of rank $n$. We construct rational curves on these moduli spaces which generalize Hecke curves on the moduli space of vector bundles. As a main result, we show that these Hecke type curves have the minimal degree among the rational curves passing through a general point of the moduli spaces. As its byproducts, we show the non-abelian Torelli theorem and compute the automorphism group of moduli spaces., Comment: 23 pages

Many emerging applications in microscale engineering rely on the fabrication of three-dimensional architectures in inorganic materials. Small-scale additive manufacturing (AM) aspires to provide flexible and facile access to these geometries. Yet, the synthesis of device-grade inorganic materials is still a key challenge towards the implementation of AM in microfabrication. Here, we present a comprehensive overview of the microstructural and mechanical properties of metals fabricated by most state-of-the-art AM methods that offer a spatial resolution $\leq$10$\mu$m. Standardized sets of samples were studied by cross-sectional electron microscopy, nanoindentation and microcompression. We show that current microscale AM techniques synthesize metals with a wide range of microstructures and elastic and plastic properties, including materials of dense and crystalline microstructure with excellent mechanical properties that compare well to those of thin-film nanocrystalline materials. The large variation in materials performance can be related to the individual microstructure, which in turn is coupled to the various physico-chemical principles exploited by the different printing methods. The study provides practical guidelines for users of small-scale additive methods and establishes a baseline for the future optimization of the properties of printed metallic objects $-$ a significant step towards the potential establishment of AM techniques in microfabrication.

In this paper, we construct proper pushforwards and flat pullbacks in Chow groups of coherent sheaf stacks over a Deligne-Mumford(DM) stack. When there is a relative semi-perfect obstruction theory for a DM-type morphism $X \to Y$, $X$ is a DM stack and $Y$ is a DM stack or a smooth Artin stack, we define a virtual pull-back as a bivariant class. This is an analogue of virtual pull-backs defined by Manolache., Comment: 49 pages, all comments are welcome

We analyse the local structure of moduli space of genus one stable quasimaps. Combining it with the p-fields theory developed in \cite{L}, we prove the splitting formula for the virtual cycle of stable quasimaps to complete intersections in $\PP^n$.

A. Zinger defined reduced Gromov-Witten (GW) invariants and proved a comparison theorem of standard and reduced genus one GW invariants for every symplectic manifold (with all dimension). H. -L. Chang and J. Li provided a proof of the comparison theorem for quintic Calabi-Yau 3-folds in algebraic geometry by taking a definition of reduced invariants as an Euler number of certain vector bundle. T. Coates and C. Manolache have defined reduced GW invariants in algebraic geometry following the idea by Vakil and Zinger and proved the comparison theorem for every Calabi-Yau threefold. In this paper, we prove the comparison theorem for every (not necessarily Calabi-Yau) complete intersection of dimension 2 or 3 in projective spaces by taking a definition of reduced GW invariants in the paper of Coates and Manolache., Comment: 27 pages

Let $(S,L_{S})$ be a polarized abelian surface, and let $M = c \cdot \pi^*L_S - \alpha \cdot \sum_{i=1}^r E_i$ be a line bundle on ${\rm Bl}_{r}(S)$, where $\pi:{\rm Bl}_{r}(S) \rightarrow S$ is the blow-up of $S$ at $r$ general points with exceptional divisors $E_{1},\dots,E_{r}$. In this paper, we provide a criterion for $k$-very ampleness of $M$. Also, we deal with the case when $S$ is an arbitrary surface of Picard number one with a numerically trivial canonical divisor., Comment: 17 pages, comments are welcome, minor changes on Question 4.4 and Theorem 4.5

We prove that the moduli spaces of rational curves of degree at most $3$ in linear sections of the Grassmannian $Gr(2,5)$ are all rational varieties. We also study their compactifications and birational geometry.

Let $X$ be a smooth projective curve with genus $g\geq3$. Let $\mathcal{N}$ be the moduli space of stable rank two vector bundles on $X$ with a fixed determinant $\mathcal{O}_X(-x)$ for $x\in X$. In this paper, as a generalization of Kiem and Castravet's works, we study the stable maps in $\mathcal{N}$ with genus $0$ and degree $3$. Let $P$ be a natural closed subvariety of $\mathcal{N}$ which parametrizes stable vector bundles with a fixed subbundle $L^{-1}(-x)$ for a line bundle $L$ on $X$. We describe the stable map space $\mathbf{M}_0(P,3)$. It turns out that the space $\mathbf{M}_0(P,3)$ consists of two irreducible components. One of them parameterizes smooth rational cubic curves and the other parameterizes the union of line and smooth conics.