Search

A wide array of sequence models are built on a framework modeled after Transformers, comprising alternating sequence mixer and channel mixer layers. This paper studies a unifying matrix mixer view of sequence mixers that can be conceptualized as a linear map on the input sequence. This framework encompasses a broad range of well-known sequence models, including the self-attention of Transformers as well as recent strong alternatives such as structured state space models (SSMs), and allows understanding downstream characteristics such as efficiency and expressivity through properties of their structured matrix class. We identify a key axis of matrix parameterizations termed sequence alignment, which increases the flexibility and performance of matrix mixers, providing insights into the strong performance of Transformers and recent SSMs such as Mamba. Furthermore, the matrix mixer framework offers a systematic approach to developing sequence mixers with desired properties, allowing us to develop several new sub-quadratic sequence models. In particular, we propose a natural bidirectional extension of the Mamba model (Hydra), parameterized as a quasiseparable matrix mixer, which demonstrates superior performance over other sequence models including Transformers on non-causal tasks. As a drop-in replacement for attention layers, Hydra outperforms BERT by 0.8 points on the GLUE benchmark and ViT by 2% Top-1 accuracy on ImageNet.

Photoacoustic tomography (PAT) is a hybrid medical imaging technique that offer high contrast and a high spatial resolution. One challenging mathematical problem associated with PAT is reconstructing the initial pressure of the wave equation from data collected at the specific surface where the detectors are positioned. The study addresses this problem when PAT is modeled by a wave equation with unknown sound speed $c$, which is a function of spatial variables, and under the assumption that both the Dirichlet and Neumann boundary values on the detector surface are measured. In practical, we introduce a novel implicit learning framework to simultaneously estimate the unknown $c$ and the reconstruction operator using only Dirichlet and Neumann boundary measurement data. The experimental results confirm the success of our proposed framework, demonstrating its ability to accurately estimate variable sound speed and the reconstruction operator in PAT.

This short article highlights the unsolved problems of magnetic reconnection in collisionless plasma. The advanced in-situ plasma measurements and simulations enabled scientists to gain a novel understanding of magnetic reconnection. Still, outstanding questions remain on the complex dynamics and structures in the diffusion region, on the cross-scale and regional couplings, on the onset of magnetic reconnection, and on the details of energetics. Future directions of the magnetic reconnection research in terms of new observations, new simulations and interdisciplinary approaches are discussed., Comment: Submitted to Space Science Reviews. This is a review paper and is an outcome of the International Space Science Institute (ISSI) Workshop on Magnetic Reconnection: Explosive Energy Conversion in Space Plasmas held at June 27 - July 1, 2022

We present a high-fidelity Mixed Reality sensor emulation framework for testing and evaluating the resilience of Unmanned Aerial Vehicles (UAVs) against false data injection (FDI) attacks. The proposed approach can be utilized to assess the impact of FDI attacks, benchmark attack detector performance, and validate the effectiveness of mitigation/reconfiguration strategies in single-UAV and UAV swarm operations. Our Mixed Reality framework leverages high-fidelity simulations of Gazebo and a Motion Capture system to emulate proprioceptive (e.g., GNSS) and exteroceptive (e.g., camera) sensor measurements in real-time. We propose an empirical approach to faithfully recreate signal characteristics such as latency and noise in these measurements. Finally, we illustrate the efficacy of our proposed framework through a Mixed Reality experiment consisting of an emulated GNSS attack on an actual UAV, which (i) demonstrates the impact of false data injection attacks on GNSS measurements and (ii) validates a mitigation strategy utilizing a distributed camera network developed in our previous work. Our open-source implementation is available at \href{https://github.com/CogniPilot/mixed\_sense}{\texttt{https://github.com/CogniPilot/mixed\_sense}}, Comment: 6 pages, 5 figures, IROS 2024

Retrieval-augmented generation supports language models to strengthen their factual groundings by providing external contexts. However, language models often face challenges when given extensive information, diminishing their effectiveness in solving questions. Context compression tackles this issue by filtering out irrelevant information, but current methods still struggle in realistic scenarios where crucial information cannot be captured with a single-step approach. To overcome this limitation, we introduce CompAct, a novel framework that employs an active strategy to condense extensive documents without losing key information. Our experiments demonstrate that CompAct brings significant improvements in both performance and compression rate on multi-hop question-answering benchmarks. CompAct flexibly operates as a cost-efficient plug-in module with various off-the-shelf retrievers or readers, achieving exceptionally high compression rates (47x)., Comment: Accepted to the main conference at EMNLP 2024

We present a review of popular simulation engines and frameworks used in reinforcement learning (RL) research, aiming to guide researchers in selecting tools for creating simulated physical environments for RL and training setups. It evaluates nine frameworks (Brax, Chrono, Gazebo, MuJoCo, ODE, PhysX, PyBullet, Webots, and Unity) based on their popularity, feature range, quality, usability, and RL capabilities. We highlight the challenges in selecting and utilizing physics engines for RL research, including the need for detailed comparisons and an understanding of each framework's capabilities. Key findings indicate MuJoCo as the leading framework due to its performance and flexibility, despite usability challenges. Unity is noted for its ease of use but lacks scalability and simulation fidelity. The study calls for further development to improve simulation engines' usability and performance and stresses the importance of transparency and reproducibility in RL research. This review contributes to the RL community by offering insights into the selection process for simulation engines, facilitating informed decision-making., Comment: 11 pages, 3 figures

Reducing scan time in Positron Emission Tomography (PET) imaging while maintaining high-quality images is crucial for minimizing patient discomfort and radiation exposure. Due to the limited size of datasets and distribution discrepancy across scanners in medical imaging, fine-tuning in a parameter-efficient and effective manner is on the rise. Motivated by the potential of Parameter-Efficient Fine-Tuning (PEFT), we aim to address these issues by effectively leveraging PEFT to improve limited data and GPU resource issues in multi-scanner setups. In this paper, we introduce PETITE, Parameter-Efficient Fine-Tuning for MultI-scanner PET to PET REconstruction that uses fewer than 1% of the parameters. To the best of our knowledge, this study is the first to systematically explore the efficacy of diverse PEFT techniques in medical imaging reconstruction tasks via prevalent encoder-decoder-type deep models. This investigation, in particular, brings intriguing insights into PETITE as we show further improvements by treating encoder and decoder separately and mixing different PEFT methods, namely, Mix-PEFT. Using multi-scanner PET datasets comprised of five different scanners, we extensively test the cross-scanner PET scan time reduction performances (i.e., a model pre-trained on one scanner is fine-tuned on a different scanner) of 21 feasible Mix-PEFT combinations to derive optimal PETITE. We show that training with less than 1% parameters using PETITE performs on par with full fine-tuning (i.e., 100% parameter)

On the setting of positive definite operators we study the near-order properties of power means such as the quasi-arithmetic mean (H\"{o}lder mean) and R\'{e}nyi power mean. We see the monotonicity of spectral geometric mean and Wasserstein mean on parameters with respect to the near-order and the near-order relationship between the spectral geometric mean and Wasserstein mean. Furthermore, the monotonicity of quasi-arithmetic mean on parameters and the convergence of R\'{e}nyi power mean to the log-Euclidean mean with respect to the near-order have been established.

The ability to communicate uncertainty, risk, and limitation is crucial for the safety of large language models. However, current evaluations of these abilities rely on simple calibration, asking whether the language generated by the model matches appropriate probabilities. Instead, evaluation of this aspect of LLM communication should focus on the behaviors of their human interlocutors: how much do they rely on what the LLM says? Here we introduce an interaction-centered evaluation framework called Rel-A.I. (pronounced "rely"}) that measures whether humans rely on LLM generations. We use this framework to study how reliance is affected by contextual features of the interaction (e.g, the knowledge domain that is being discussed), or the use of greetings communicating warmth or competence (e.g., "I'm happy to help!"). We find that contextual characteristics significantly affect human reliance behavior. For example, people rely 10% more on LMs when responding to questions involving calculations and rely 30% more on LMs that are perceived as more competent. Our results show that calibration and language quality alone are insufficient in evaluating the risks of human-LM interactions, and illustrate the need to consider features of the interactional context., Comment: Preprint

This working note outlines our participation in the retrieval task at CLEF 2024. We highlight the considerable gap between studying retrieval performance on static knowledge documents and understanding performance in real-world environments. Therefore, Addressing these discrepancies and measuring the temporal persistence of IR systems is crucial. By investigating the LongEval benchmark, specifically designed for such dynamic environments, our findings demonstrate the effectiveness of a listwise reranking approach, which proficiently handles inaccuracies induced by temporal distribution shifts. Among listwise rerankers, our findings show that ListT5, which effectively mitigates the positional bias problem by adopting the Fusion-in-Decoder architecture, is especially effective, and more so, as temporal drift increases, on the test-long subset., Comment: Accepted at CLEF 2024 LongEval track. Abstract revised: its first two (background) sentences were too similar to an earlier paper arXiv:2305.18952

Analog front-end design heavily relies on specialized human expertise and costly trial-and-error simulations, which motivated many prior works on analog design automation. However, efficient and effective exploration of the vast and complex design space remains constrained by the time-consuming nature of SPICE simulations, making effective design automation a challenging endeavor. In this paper, we introduce INSIGHT, a GPU-powered, technology-agnostic, effective universal neural simulator in the analog front-end design automation loop. INSIGHT accurately predicts the performance metrics of analog circuits across various technologies with just a few microseconds of inference time. Notably, its autoregressive capabilities enable INSIGHT to accurately predict simulation-costly critical transient specifications leveraging less expensive performance metric information. The low cost and high fidelity feature make INSIGHT a good substitute for standard simulators in analog front-end optimization frameworks. INSIGHT is compatible with any optimization framework, facilitating enhanced design space exploration for sample efficiency through sophisticated offline learning and adaptation techniques. Our experiments demonstrate that INSIGHT-M, a model-based batch reinforcement learning sizing framework with INSIGHT as the accurate surrogate, only requires < 20 real-time simulations with 100-1000x lower simulation costs and significant speedup over existing sizing methods.

In this paper, we propose an experiment, LANSCE-mQ, aiming to detect fractionally charged and millicharged particles (mCP) using an 800 MeV proton beam fixed target at the Los Alamos Neutron Science Center (LANSCE) facility. This search can shed new light on numerous fundamental questions, including charge quantization, the predictions of string theories and grand unification theories, the gauge symmetry of the Standard Model, dark sector models, and the tests of cosmic reheating. We propose to install two-layer scintillation detectors made of plastic (such as EJ-200) or CeBr3 to search for mCPs. Dedicated Geant4 detector simulations and in situ measurements have been conducted to obtain a preliminary determination of the background rate. The dominant backgrounds are beam-induced neutrons and coincident dark current signals from the photomultiplier tubes, while beam-induced gammas and cosmic muons are subdominant. We determined that LANSCE-mQ, the dedicated mCP experiment, has the leading mCP sensitivity for mass between ~ 1 MeV to 300 MeV., Comment: 8 pages, 8 figures

AMoRE searches for the signature of neutrinoless double beta decay of $^{100}$Mo with a 100 kg sample of enriched $^{100}$Mo. Scintillating molybdate crystals coupled with a metallic magnetic calorimeter operate at milli-Kelvin temperatures to measure the energy of electrons emitted in the decay. As a demonstration of the full-scale AMoRE, we conducted AMoRE-I, a pre-experiment with 18 molybdate crystals, at the Yangyang Underground Laboratory for over two years. The exposure was 8.02 kg$\cdot$year (or 3.89 kg$_{\mathrm{^{100}Mo}}\cdot$year) and the total background rate near the Q-value was 0.025 $\pm$ 0.002 counts/keV/kg/year. We observed no indication of $0\nu\beta\beta$ decay and report a new lower limit of the half-life of $^{100}$Mo $0\nu\beta\beta$ decay as $ T^{0\nu}_{1/2}>3.0\times10^{24}~\mathrm{years}$ at 90\% confidence level. The effective Majorana mass limit range is $m_{\beta\beta}<$(210--610) meV using nuclear matrix elements estimated in the framework of different models, including the recent shell model calculations., Comment: 8 pages, 5 figures

In neuroimaging, generally, brain CT is more cost-effective and accessible imaging option compared to MRI. Nevertheless, CT exhibits inferior soft-tissue contrast and higher noise levels, yielding less precise structural clarity. In response, leveraging more readily available CT to construct its counterpart MRI, namely, medical image-to-image translation (I2I), serves as a promising solution. Particularly, while diffusion models (DMs) have recently risen as a powerhouse, they also come with a few practical caveats for medical I2I. First, DMs' inherent stochasticity from random noise sampling cannot guarantee consistent MRI generation that faithfully reflects its CT. Second, for 3D volumetric images which are prevalent in medical imaging, naively using 2D DMs leads to slice inconsistency, e.g., abnormal structural and brightness changes. While 3D DMs do exist, significant training costs and data dependency bring hesitation. As a solution, we propose novel style key conditioning (SKC) and inter-slice trajectory alignment (ISTA) sampling for the 2D Brownian bridge diffusion model. Specifically, SKC ensures a consistent imaging style (e.g., contrast) across slices, and ISTA interconnects the independent sampling of each slice, deterministically achieving style and shape consistent 3D CT-to-MRI translation. To the best of our knowledge, this study is the first to achieve high-quality 3D medical I2I based only on a 2D DM with no extra architectural models. Our experimental results show superior 3D medical I2I than existing 2D and 3D baselines, using in-house CT-MRI dataset and BraTS2023 FLAIR-T1 MRI dataset., Comment: 13 pages, 7 figures, Early accepted at Medical Image Computing and Computer Assisted Intervention (MICCAI) 2024

Graph classification benchmarks, vital for assessing and developing graph neural networks (GNNs), have recently been scrutinized, as simple methods like MLPs have demonstrated comparable performance. This leads to an important question: Do these benchmarks effectively distinguish the advancements of GNNs over other methodologies? If so, how do we quantitatively measure this effectiveness? In response, we first propose an empirical protocol based on a fair benchmarking framework to investigate the performance discrepancy between simple methods and GNNs. We further propose a novel metric to quantify the dataset effectiveness by considering both dataset complexity and model performance. To the best of our knowledge, our work is the first to thoroughly study and provide an explicit definition for dataset effectiveness in the graph learning area. Through testing across 16 real-world datasets, we found our metric to align with existing studies and intuitive assumptions. Finally, we explore the causes behind the low effectiveness of certain datasets by investigating the correlation between intrinsic graph properties and class labels, and we developed a novel technique supporting the correlation-controllable synthetic dataset generation. Our findings shed light on the current understanding of benchmark datasets, and our new platform could fuel the future evolution of graph classification benchmarks.

In this work, we introduce a novel tracker designed for online multiple object tracking with a focus on being simple, while being effective. we provide multiple feature modules each of which stands for a particular appearance information. By integrating distinct appearance features, including clothing color, style, and target direction, alongside a ReID network for robust embedding extraction, our tracker significantly enhances online tracking accuracy. Additionally, we propose the incorporation of a stronger detector and also provide an advanced post processing methods that further elevate the tracker's performance. During real time operation, we establish measurement to track associated distance function which includes the IoU, direction, color, style, and ReID features similarity information, where each metric is calculated separately. With the design of our feature related distance function, it is possible to track objects through longer period of occlusions, while keeping the number of identity switches comparatively low. Extensive experimental evaluation demonstrates notable improvement in tracking accuracy and reliability, as evidenced by reduced identity switches and enhanced occlusion handling. These advancements not only contribute to the state of the art in object tracking but also open new avenues for future research and practical applications demanding high precision and reliability.

Recent advancements in Large Language Models (LLMs) have significantly improved their performance across various Natural Language Processing (NLP) tasks. However, LLMs still struggle with generating non-factual responses due to limitations in their parametric memory. Retrieval-Augmented Generation (RAG) systems address this issue by incorporating external knowledge with a retrieval module. Despite their successes, however, current RAG systems face challenges with retrieval failures and the limited ability of LLMs to filter out irrelevant information. Therefore, in this work, we propose DSLR (Document Refinement with Sentence-Level Re-ranking and Reconstruction), an unsupervised framework that decomposes retrieved documents into sentences, filters out irrelevant sentences, and reconstructs them again into coherent passages. We experimentally validate DSLR on multiple open-domain QA datasets and the results demonstrate that DSLR significantly enhances the RAG performance over conventional fixed-size passage. Furthermore, our DSLR enhances performance in specific, yet realistic scenarios without the need for additional training, providing an effective and efficient solution for refining retrieved documents in RAG systems., Comment: 20 pages

This paper presents a cooperative multi-agent deep reinforcement learning (MADRL) approach for unmmaned aerial vehicle (UAV)-aided mobile edge computing (MEC) networks. An UAV with computing capability can provide task offlaoding services to ground internet-of-things devices (IDs). With partial observation of the entire network state, the UAV and the IDs individually determine their MEC strategies, i.e., UAV trajectory, resource allocation, and task offloading policy. This requires joint optimization of decision-making process and coordination strategies among the UAV and the IDs. To address this difficulty, the proposed cooperative MADRL approach computes two types of action variables, namely message action and solution action, each of which is generated by dedicated actor neural networks (NNs). As a result, each agent can automatically encapsulate its coordination messages to enhance the MEC performance in the decentralized manner. The proposed actor structure is designed based on graph attention networks such that operations are possible regardless of the number of IDs. A scalable training algorithm is also proposed to train a group of NNs for arbitrary network configurations. Numerical results demonstrate the superiority of the proposed cooperative MADRL approach over conventional methods., Comment: 13 pages, 6 figures

Neural rendering-based urban scene reconstruction methods commonly rely on images collected from driving vehicles with cameras facing and moving forward. Although these methods can successfully synthesize from views similar to training camera trajectory, directing the novel view outside the training camera distribution does not guarantee on-par performance. In this paper, we tackle the Extrapolated View Synthesis (EVS) problem by evaluating the reconstructions on views such as looking left, right or downwards with respect to training camera distributions. To improve rendering quality for EVS, we initialize our model by constructing dense LiDAR map, and propose to leverage prior scene knowledge such as surface normal estimator and large-scale diffusion model. Qualitative and quantitative comparisons demonstrate the effectiveness of our methods on EVS. To the best of our knowledge, we are the first to address the EVS problem in urban scene reconstruction. Link to our project page: https://vegs3d.github.io/., Comment: The first two authors contributed equally. Project Page: https://vegs3d.github.io/

Sign language, essential for the deaf and hard-of-hearing, presents unique challenges in translation and production due to its multimodal nature and the inherent ambiguity in mapping sign language motion to spoken language words. Previous methods often rely on gloss annotations, requiring time-intensive labor and specialized expertise in sign language. Gloss-free methods have emerged to address these limitations, but they often depend on external sign language data or dictionaries, failing to completely eliminate the need for gloss annotations. There is a clear demand for a comprehensive approach that can supplant gloss annotations and be utilized for both Sign Language Translation (SLT) and Sign Language Production (SLP). We introduce Universal Gloss-level Representation (UniGloR), a unified and self-supervised solution for both SLT and SLP, trained on multiple datasets including PHOENIX14T, How2Sign, and NIASL2021. Our results demonstrate UniGloR's effectiveness in the translation and production tasks. We further report an encouraging result for the Sign Language Recognition (SLR) on previously unseen data. Our study suggests that self-supervised learning can be made in a unified manner, paving the way for innovative and practical applications in future research., Comment: 14 pages, 5 figures

Scaling magnets down to where quantum size effects become prominent triggers quantum tunneling of magnetization (QTM), profoundly influencing magnetization dynamics. Measuring magnetization switching in an Fe atomic chain under a carefully tuned transverse magnetic field, we observe a non-monotonic variation of magnetization lifetimes around a level crossing, known as the diabolic point (DP). Near DPs, local environment effects causing QTM are efficiently suppressed, enhancing lifetimes by three orders of magnitude. Adjusting interatomic interactions further facilitates multiple DPs. Our study provides a deeper understanding of quantum dynamics near DPs and enhances our ability to engineer a quantum magnet., Comment: Main text and Supplementary

Using new continuum and molecular line data from the ALMA Three-millimeter Observations of Massive Star-forming Regions (ATOMS) survey and archival VLA, 4.86 GHz data, we present direct observational evidence of hierarchical triggering relating three epochs of massive star formation in a ring-like H II region, G24.47+0.49. We find from radio flux analysis that it is excited by a massive star(s) of spectral type O8.5V-O8V from the first epoch of star formation. The swept-up ionized ring structure shows evidence of secondary collapse, and within this ring a burst of massive star formation is observed in different evolutionary phases, which constitutes the second epoch. ATOMS spectral line (e.g., HCO$^+$(1-0)) observations reveal an outer concentric molecular gas ring expanding at a velocity of $\sim$ 9 $\rm km\,s^{-1}$, constituting the direct and unambiguous detection of an expanding molecular ring. It harbors twelve dense molecular cores with surface mass density greater than 0.05 $\rm g\,cm^{-2}$, a threshold typical of massive star formation. Half of them are found to be subvirial, and thus in gravitational collapse, making them third epoch of potential massive star-forming sites., Comment: 18 pages, 7 figures, accepted for publication in The Astrophysical Journal Letters

Image dehazing, addressing atmospheric interference like fog and haze, remains a pervasive challenge crucial for robust vision applications such as surveillance and remote sensing under adverse visibility. While various methodologies have evolved from early works predicting transmission matrix and atmospheric light features to deep learning and dehazing networks, they innately prioritize dehazing quality metrics, neglecting the need for real-time applicability in time-sensitive domains like autonomous driving. This work introduces FALCON (Frequency Adjoint Link with CONtinuous density mask), a single-image dehazing system achieving state-of-the-art performance on both quality and speed. Particularly, we develop a novel bottleneck module, namely, Frequency Adjoint Link, operating in the frequency space to globally expand the receptive field with minimal growth in network size. Further, we leverage the underlying haze distribution based on the atmospheric scattering model via a Continuous Density Mask (CDM) which serves as a continuous-valued mask input prior and a differentiable auxiliary loss. Comprehensive experiments involving multiple state-of-the-art methods and ablation analysis demonstrate FALCON's exceptional performance in both dehazing quality and speed (i.e., >$180 frames-per-second), quantified by metrics such as FPS, PSNR, and SSIM.

We present a maximum entropy inverse reinforcement learning (IRL) approach for improving the sample quality of diffusion generative models, especially when the number of generation time steps is small. Similar to how IRL trains a policy based on the reward function learned from expert demonstrations, we train (or fine-tune) a diffusion model using the log probability density estimated from training data. Since we employ an energy-based model (EBM) to represent the log density, our approach boils down to the joint training of a diffusion model and an EBM. Our IRL formulation, named Diffusion by Maximum Entropy IRL (DxMI), is a minimax problem that reaches equilibrium when both models converge to the data distribution. The entropy maximization plays a key role in DxMI, facilitating the exploration of the diffusion model and ensuring the convergence of the EBM. We also propose Diffusion by Dynamic Programming (DxDP), a novel reinforcement learning algorithm for diffusion models, as a subroutine in DxMI. DxDP makes the diffusion model update in DxMI efficient by transforming the original problem into an optimal control formulation where value functions replace back-propagation in time. Our empirical studies show that diffusion models fine-tuned using DxMI can generate high-quality samples in as few as 4 and 10 steps. Additionally, DxMI enables the training of an EBM without MCMC, stabilizing EBM training dynamics and enhancing anomaly detection performance., Comment: Code is released at https://github.com/swyoon/Diffusion-by-MaxEntIRL

Despite recent advancements in vision-language models, their performance remains suboptimal on images from non-western cultures due to underrepresentation in training datasets. Various benchmarks have been proposed to test models' cultural inclusivity, but they have limited coverage of cultures and do not adequately assess cultural diversity across universal as well as culture-specific local concepts. To address these limitations, we introduce the GlobalRG benchmark, comprising two challenging tasks: retrieval across universals and cultural visual grounding. The former task entails retrieving culturally diverse images for universal concepts from 50 countries, while the latter aims at grounding culture-specific concepts within images from 15 countries. Our evaluation across a wide range of models reveals that the performance varies significantly across cultures -- underscoring the necessity for enhancing multicultural understanding in vision-language models., Comment: Under peer review

Large Language Models (LLMs) exhibit strong generalization capabilities to novel tasks when prompted with language instructions and in-context demos. Since this ability sensitively depends on the quality of prompts, various methods have been explored to automate the instruction design. While these methods demonstrated promising results, they also restricted the searched prompt to one instruction. Such simplification significantly limits their capacity, as a single demo-free instruction might not be able to cover the entire complex problem space of the targeted task. To alleviate this issue, we adopt the Mixture-of-Expert paradigm and divide the problem space into a set of sub-regions; Each sub-region is governed by a specialized expert, equipped with both an instruction and a set of demos. A two-phase process is developed to construct the specialized expert for each region: (1) demo assignment: Inspired by the theoretical connection between in-context learning and kernel regression, we group demos into experts based on their semantic similarity; (2) instruction assignment: A region-based joint search of an instruction per expert complements the demos assigned to it, yielding a synergistic effect. The resulting method, codenamed Mixture-of-Prompts (MoP), achieves an average win rate of 81% against prior arts across several major benchmarks., Comment: ICML 2024. code available at https://github.com/ruocwang/mixture-of-prompts

While shape optimization using isogeometric shells exhibits appealing features by integrating design geometries and analysis models, challenges arise when addressing computer-aided design (CAD) geometries comprised of multiple non-uniform rational B-splines (NURBS) patches, which are common in practice. The intractability stems from surface intersections within these CAD models. In this paper, we develop an approach for shape optimization of non-matching isogeometric shells incorporating intersection movement. Separately parametrized NURBS surfaces are modeled using Kirchhoff--Love shell theory and coupled using a penalty-based formulation. The optimization scheme allows shell patches to move without preserving relative location with other members during the shape optimization. This flexibility is achieved through an implicit state function, and analytical sensitivities are derived for the relative movement of shell patches. The introduction of differentiable intersections expands the design space and overcomes challenges associated with large mesh distortion, particularly when optimal shapes involve significant movement of patch intersections in physical space. Throughout optimization iterations, all members within the shell structures maintain the NURBS geometry representation, enabling efficient integration of analysis and design models. The optimization approach leverages the multilevel design concept by selecting a refined model for accurate analysis from a coarse design model while maintaining the same geometry. We adopt several example problems to verify the effectiveness of the proposed scheme and demonstrate its applicability to the optimization of the internal stiffeners of an aircraft wing., Comment: 41 pages, 18 figures

Large-scale generative models, such as text-to-image diffusion models, have garnered widespread attention across diverse domains due to their creative and high-fidelity image generation. Nonetheless, existing large-scale diffusion models are confined to generating images of up to 1K resolution, which is far from meeting the demands of contemporary commercial applications. Directly sampling higher-resolution images often yields results marred by artifacts such as object repetition and distorted shapes. Addressing the aforementioned issues typically necessitates training or fine-tuning models on higher-resolution datasets. However, this poses a formidable challenge due to the difficulty in collecting large-scale high-resolution images and substantial computational resources. While several preceding works have proposed alternatives to bypass the cumbersome training process, they often fail to produce convincing results. In this work, we probe the generative ability of diffusion models at higher resolution beyond their original capability and propose a novel progressive approach that fully utilizes generated low-resolution images to guide the generation of higher-resolution images. Our method obviates the need for additional training or fine-tuning which significantly lowers the burden of computational costs. Extensive experiments and results validate the efficiency and efficacy of our method. Project page: https://yhyun225.github.io/DiffuseHigh/, Comment: Project page: https://yhyun225.github.io/DiffuseHigh/

Cable-Driven Continuum Manipulators (CDCMs) enable scar-free procedures but face limitations in workspace and control accuracy due to hysteresis. We introduce an extensible CDCM with a Semi-active Mechanism (SAM) and develop a real-time hysteresis compensation control algorithm using a Temporal Convolutional Network (TCN) based on data collected from fiducial markers and RGBD sensing. Performance validation shows the proposed controller significantly reduces hysteresis by up to 69.5% in random trajectory tracking test and approximately 26% in the box pointing task. The SAM mechanism enables access to various lesions without damaging surrounding tissues. The proposed controller with TCN-based compensation effectively predicts hysteresis behavior and minimizes position and joint angle errors in real-time, which has the potential to enhance surgical task performance., Comment: 22 pages, 19 figures, 9 tables

Information retrieval models that aim to search for the documents relevant to the given query have shown many successes, which have been applied to diverse tasks. However, the query provided by the user is oftentimes very short, which challenges the retrievers to correctly fetch relevant documents. To tackle this, existing studies have proposed expanding the query with a couple of additional (user-related) features related to the query. Yet, they may be suboptimal to effectively augment the query, though there is plenty of information available to augment it in a relational database. Motivated by this, we present a novel retrieval framework called Database-Augmented Query representation (DAQu), which augments the original query with various (query-related) metadata across multiple tables. In addition, as the number of features in the metadata can be very large and there is no order among them, we encode them with our graph-based set encoding strategy, which considers hierarchies of features in the database without order. We validate DAQu in diverse retrieval scenarios that can incorporate metadata from the relational database, demonstrating that ours significantly enhances overall retrieval performance, compared to existing query augmentation methods.

The context window within a transformer provides a form of active memory for the current task, which can be useful for few-shot learning and conditional generation, both which depend heavily on previous context tokens. However, as the context length grows, the computational cost increases quadratically. Recent works have shown that saving a few initial tokens along with a fixed-sized sliding window leads to stable streaming generation with linear complexity in transformer-based Large Language Models (LLMs). However, they make suboptimal use of the fixed window by naively evicting all tokens unconditionally from the key-value (KV) cache once they reach the end of the window, resulting in tokens being forgotten and no longer able to affect subsequent predictions. To overcome this limitation, we propose a novel mechanism for storing longer sliding window contexts with the same total cache size by keeping separate cascading sub-cache buffers whereby each subsequent buffer conditionally accepts a fraction of the relatively more important tokens evicted from the previous buffer. Our method results in a dynamic KV cache that can store tokens from the more distant past than a fixed, static sliding window approach. Our experiments show improvements of 5.6% on long context generation (LongBench), 1.2% in streaming perplexity (PG19), and 0.6% in language understanding (MMLU STEM) using LLMs given the same fixed cache size. Additionally, we provide an efficient implementation that improves the KV cache latency from 1.33ms per caching operation to 0.54ms, a 59% speedup over previous work.

Much research in machine learning involves finding appropriate inductive biases (e.g. convolutional neural networks, momentum-based optimizers, transformers) to promote generalization on tasks. However, quantification of the amount of inductive bias associated with these architectures and hyperparameters has been limited. We propose a novel method for efficiently computing the inductive bias required for generalization on a task with a fixed training data budget; formally, this corresponds to the amount of information required to specify well-generalizing models within a specific hypothesis space of models. Our approach involves modeling the loss distribution of random hypotheses drawn from a hypothesis space to estimate the required inductive bias for a task relative to these hypotheses. Unlike prior work, our method provides a direct estimate of inductive bias without using bounds and is applicable to diverse hypothesis spaces. Moreover, we derive approximation error bounds for our estimation approach in terms of the number of sampled hypotheses. Consistent with prior results, our empirical results demonstrate that higher dimensional tasks require greater inductive bias. We show that relative to other expressive model classes, neural networks as a model class encode large amounts of inductive bias. Furthermore, our measure quantifies the relative difference in inductive bias between different neural network architectures. Our proposed inductive bias metric provides an information-theoretic interpretation of the benefits of specific model architectures for certain tasks and provides a quantitative guide to developing tasks requiring greater inductive bias, thereby encouraging the development of more powerful inductive biases., Comment: Published at IJCAI 2024

We report on ALMA observations of polarized dust emission at 1.2 mm from NGC6334I, a source known for its significant flux outbursts. Between five months, our data show no substantial change in total intensity and a modest 8\% variation in linear polarization, suggesting a phase of stability or the conclusion of the outburst. The magnetic field, inferred from this polarized emission, displays a predominantly radial pattern from North-West to South-East with intricate disturbances across major cores, hinting at spiral structures. Energy analysis of CS$(J=5 \rightarrow 4)$ emission yields an outflow energy of approximately $3.5\times10^{45}$ ergs, aligning with previous interferometric studies. Utilizing the Davis-Chandrasekhar-Fermi method, we determined magnetic field strengths ranging from 1 to 11 mG, averaging at 1.9 mG. This average increases to 4 $\pm 1$ mG when incorporating Zeeman measurements. Comparative analyses using gravitational, thermal, and kinetic energy maps reveal that magnetic energy is significantly weaker, possibly explaining the observed field morphology. We also find that the energy in the outflows and the expanding cometary {\HII} region is also larger than the magnetic energy, suggesting that protostellar feedback maybe the dominant driver behind the injection of turbulence in NGC6334I at the scales sampled by our data. The gas in NGC6334I predominantly exhibits supersonic and trans-Alfvenic conditions, transitioning towards a super-Alfvenic regime, underscoring a diminished influence of the magnetic field with increasing gas density. These observations are in agreement with prior polarization studies at 220 GHz, enriching our understanding of the dynamic processes in high-mass star-forming regions., Comment: Accepted for Publication at the Astrophysical Journal

The short-wave infrared (SWIR) wavelength, especially 1.55 um, has attracted significant attention in various areas such as high-speed optical communication and LiDAR systems. Avalanche photodiodes (APDs) are a critical component as a receiver in these systems due to their internal gain which enhances the system performance. Silicon-based APDs are promising since they are CMOS compatible, but they are limited in detecting 1.55 um light detection. This study proposes a p-type Si on n-type GaAs0.51Sb0.49 (GaAsSb) lattice matched to InP substrates heterojunction formed using a grafting technique for future GaAsSb/Si APD technology. A p+Si nanomembrane is transferred onto the GaAsSb/AlInAs/InP substrate, with an ultrathin ALD-Al2O3 oxide at the interface, which behaves as both double-side passivation and quantum tunneling layers. The devices exhibit excellent surface morphology and interface quality, confirmed by atomic force microscope (AFM) and transmission electron microscope (TEM). Also, the current-voltage (I-V) of the p+Si/n-GaAsSb heterojunction shows ideal rectifying characteristics with an ideality factor of 1.15. The I-V tests across multiple devices confirm high consistency and yield. Furthermore, the X-ray photoelectron spectroscopy (XPS) measurement reveals that GaAsSb and Si are found to have type-II band alignment with a conduction band offset of 50 meV which is favorable for the high-bandwidth APD application. The demonstration of the GaAsSb/Si heterojunction highlights the potential to advance current SWIR PD technologies., Comment: 14 pages, 6 figures

Recent advancements in Large Language Models (LLMs) have significantly enhanced the capabilities of conversational agents, making them applicable to various fields (e.g., education). Despite their progress, the evaluation of the agents often overlooks the complexities of real-world conversations, such as real-time interactions, multi-party dialogues, and extended contextual dependencies. To bridge this gap, we introduce DialSim, a real-time dialogue simulator. In this simulator, an agent is assigned the role of a character from popular TV shows, requiring it to respond to spontaneous questions using past dialogue information and to distinguish between known and unknown information. Key features of DialSim include evaluating the agent's ability to respond within a reasonable time limit, handling long-term multi-party dialogues, and testing the agent's performance under randomized questioning with a diverse and high-quality question-answer dataset. We utilized this simulator to evaluate the latest conversational agents and analyze their limitations. Our experiments highlight both the strengths and weaknesses of these agents, providing valuable insights for future improvements in the field of conversational AI. DialSim is available at https://dialsim.github.io/.

A heat modulator is proposed based on a voltage-biased Aharonov-Bohm interferometer. Once an electrical bias is applied, Peltier effects give rise to a flow of heat that can be modulated by a magnetic flux. We determine the corresponding temperature changes using a simple thermal model. Our calculations demonstrate that the modulated temperature difference can be as large as 80 mK at base temperature about 600 mK with relative temperature variations reaching 10\%. Our model also predicts, quite generally, the emergence of spin-polarized heat flows without any ferromagnetic contacts, if Rashba spin-orbit interaction is combined with the applied magnetic flux, which potentially paves the way towards caloritronic information processing., Comment: 8 pages, 4 figures

Group dance, a sub-genre characterized by intricate motions made by a cohort of performers in tight synchronization, has a longstanding and culturally significant history and, in modern forms such as cheerleading, a broad base of current adherents. However, despite its popularity, learning group dance routines remains challenging. Based on the prior success of interactive systems to support individual dance learning, this paper argues that group dance settings are fertile ground for augmentation by interactive aids. To better understand these design opportunities, this paper presents a sequence of user-centered studies of and with amateur cheerleading troupes, spanning from the formative (interviews, observations) through the generative (an ideation workshop) to concept validation (technology probes and speed dating). The outcomes are a nuanced understanding of the lived practice of group dance learning, a set of interactive concepts to support those practices, and design directions derived from validating the proposed concepts. Through this empirical work, we expand the design space of interactive dance practice systems from the established context of single-user practice (primarily focused on gesture recognition) to a multi-user, group-based scenario focused on feedback and communication., Comment: 20 pages, 10 figures, 1 table, to be published in the proceedings of the ACM Designing Interactive Systems Conference, 2024, (DIS '24)

3D reconstruction from multi-view images is one of the fundamental challenges in computer vision and graphics. Recently, 3D Gaussian Splatting (3DGS) has emerged as a promising technique capable of real-time rendering with high-quality 3D reconstruction. This method utilizes 3D Gaussian representation and tile-based splatting techniques, bypassing the expensive neural field querying. Despite its potential, 3DGS encounters challenges, including needle-like artifacts, suboptimal geometries, and inaccurate normals, due to the Gaussians converging into anisotropic Gaussians with one dominant variance. We propose using effective rank analysis to examine the shape statistics of 3D Gaussian primitives, and identify the Gaussians indeed converge into needle-like shapes with the effective rank 1. To address this, we introduce effective rank as a regularization, which constrains the structure of the Gaussians. Our new regularization method enhances normal and geometry reconstruction while reducing needle-like artifacts. The approach can be integrated as an add-on module to other 3DGS variants, improving their quality without compromising visual fidelity., Comment: project page: https://junhahyung.github.io/erankgs.github.io

This paper investigates the potential of LLM-based conversational agents (CAs) to enhance critical reflection and mitigate design fixation in group design work. By challenging AI-generated recommendations and prevailing group opinions, these agents address issues such as groupthink and promote a more dynamic and inclusive design process. Key design considerations include optimizing intervention timing, ensuring clarity in counterarguments, and balancing critical thinking with designers' satisfaction. CAs can also adapt to various roles, supporting individual and collective reflection. Our work aligns with the "Death of the Design Researcher?" workshop's goals, emphasizing the transformative potential of generative AI in reshaping design practices and promoting ethical considerations. By exploring innovative uses of generative AI in group design contexts, we aim to stimulate discussion and open new pathways for future research and development, ultimately contributing to practical tools and resources for design researchers., Comment: 7 pages, 2 figures, DIS2024 Workshop on 'Death of Design Researcher'

To achieve lifelong human-agent interaction, dialogue agents need to constantly memorize perceived information and properly retrieve it for response generation (RG). While prior work focuses on getting rid of outdated memories to improve retrieval quality, we argue that such memories provide rich, important contextual cues for RG (e.g., changes in user behaviors) in long-term conversations. We present Theanine, a framework for LLM-based lifelong dialogue agents. Theanine discards memory removal and manages large-scale memories by linking them based on their temporal and cause-effect relation. Enabled by this linking structure, Theanine augments RG with memory timelines - series of memories representing the evolution or causality of relevant past events. Along with Theanine, we introduce TeaFarm, a counterfactual-driven evaluation scheme, addressing the limitation of G-Eval and human efforts in measuring memory-augmented dialogue agents. A supplementary video for Theanine and data for TeaFarm are at https://huggingface.co/spaces/ResearcherScholar/Theanine., Comment: Work in Progress

Instruction tuning, or supervised finetuning on extensive task-specific data, is necessary for Large Vision-Language Models (LVLMs) to generalize well across a broad range of vision-language (VL) tasks. However, training on large VL datasets can become prohibitively expensive. In this work, we introduce COINCIDE, an effective and scalable data selection technique that uses a small model as a reference model to select visual instruction tuning data for efficient finetuning of a target LVLM, focusing on diversity and transferability. Specifically, we cluster the training data using internal activations from a small model, which identifies VL concept-skill compositions needed by a target LVLM. We then sample data from these diverse clusters by considering their density and transferability, or the ability to transfer well to other concept-skill compositions. This approach ensures the diversity of these compositions, which is vital for LVLM generalization. Extensive experiments demonstrate that COINCIDE achieves superior performance and data selection efficiency against 8 strong baselines on two distinct datasets: LLaVA-1.5 and Vision-Flan. Using only 20% of the LLaVA-1.5 dataset, COINCIDE achieves performance comparable to the LVLM finetuned on the whole dataset, with 70% reduction of the wall-clock running time. On the Vision-Flan dataset, our method achieves superior results with only 16.7% of the training data., Comment: EMNLP 2024

We investigated the nature of the anomalies appearing in four microlensing events KMT-2020-BLG-0757, KMT-2022-BLG-0732, KMT-2022-BLG-1787, and KMT-2022-BLG-1852. The light curves of these events commonly exhibit initial bumps followed by subsequent troughs that extend across a substantial portion of the light curves. We performed thorough modeling of the anomalies to elucidate their characteristics. Despite their prolonged durations, which differ from the usual brief anomalies observed in typical planetary events, our analysis revealed that each anomaly in these events originated from a planetary companion located within the Einstein ring of the primary star. It was found that the initial bump arouse when the source star crossed one of the planetary caustics, while the subsequent trough feature occurred as the source traversed the region of minor image perturbations lying between the pair of planetary caustics. The estimated masses of the host and planet, their mass ratios, and the distance to the discovered planetary systems are $(M_{\rm host}/M_\odot, M_{\rm planet}/M_{\rm J}, q/10^{-3}, \dl/{\rm kpc}) = (0.58^{+0.33}_{-0.30}, 10.71^{+6.17}_{-5.61}, 17.61\pm 2.25,6.67^{+0.93}_{-1.30})$ for KMT-2020-BLG-0757, $(0.53^{+0.31}_{-0.31}, 1.12^{+0.65}_{-0.65}, 2.01 \pm 0.07, 6.66^{+1.19}_{-1.84})$ for KMT-2022-BLG-0732, $(0.42^{+0.32}_{-0.23}, 6.64^{+4.98}_{-3.64}, 15.07\pm 0.86, 7.55^{+0.89}_{-1.30})$ for KMT-2022-BLG-1787, and $(0.32^{+0.34}_{-0.19}, 4.98^{+5.42}_{-2.94}, 8.74\pm 0.49, 6.27^{+0.90}_{-1.15})$ for KMT-2022-BLG-1852. These parameters indicate that all the planets are giants with masses exceeding the mass of Jupiter in our solar system and the hosts are low-mass stars with masses substantially less massive than the Sun., Comment: 10 pages, 12 figures, 7 tables

In modern large language models (LLMs), increasing the context length is crucial for improving comprehension and coherence in long-context, multi-modal, and retrieval-augmented language generation. While many recent transformer models attempt to extend their context length over a million tokens, they remain impractical due to the quadratic time and space complexities. Although recent works on linear and sparse attention mechanisms can achieve this goal, their real-world applicability is often limited by the need to re-train from scratch and significantly worse performance. In response, we propose a novel approach, Hierarchically Pruned Attention (HiP), which reduces the time complexity of the attention mechanism to $O(T \log T)$ and the space complexity to $O(T)$, where $T$ is the sequence length. We notice a pattern in the attention scores of pretrained LLMs where tokens close together tend to have similar scores, which we call ``attention locality''. Based on this observation, we utilize a novel tree-search-like algorithm that estimates the top-$k$ key tokens for a given query on the fly, which is mathematically guaranteed to have better performance than random attention pruning. In addition to improving the time complexity of the attention mechanism, we further optimize GPU memory usage by implementing KV cache offloading, which stores only $O(\log T)$ tokens on the GPU while maintaining similar decoding throughput. Experiments on benchmarks show that HiP, with its training-free nature, significantly reduces both prefill and decoding latencies, as well as memory usage, while maintaining high-quality generation with minimal degradation. HiP enables pretrained LLMs to scale up to millions of tokens on commodity GPUs, potentially unlocking long-context LLM applications previously deemed infeasible., Comment: 44 pages

Low-rank adaption (LoRA) is a prominent method that adds a small number of learnable parameters to the frozen pre-trained weights for parameter-efficient fine-tuning. Prompted by the question, ``Can we make its representation enough with LoRA weights solely at the final phase of finetuning without the pre-trained weights?'' In this work, we introduce Progressive Compression LoRA~(PC-LoRA), which utilizes low-rank adaptation (LoRA) to simultaneously perform model compression and fine-tuning. The PC-LoRA method gradually removes the pre-trained weights during the training process, eventually leaving only the low-rank adapters in the end. Thus, these low-rank adapters replace the whole pre-trained weights, achieving the goals of compression and fine-tuning at the same time. Empirical analysis across various models demonstrates that PC-LoRA achieves parameter and FLOPs compression rates of 94.36%/89.1% for vision models, e.g., ViT-B, and 93.42%/84.2% parameters and FLOPs compressions for language models, e.g., BERT., Comment: Accepted at T4V@CVPR

In the rapidly evolving field of healthcare, the integration of artificial intelligence (AI) has become a pivotal component in the automation of clinical workflows, ushering in a new era of efficiency and accuracy. This study focuses on the transformative capabilities of the fine-tuned KoELECTRA model in comparison to the GPT-4 model, aiming to facilitate automated information extraction from thyroid operation narratives. The current research landscape is dominated by traditional methods heavily reliant on regular expressions, which often face challenges in processing free-style text formats containing critical details of operation records, including frozen biopsy reports. Addressing this, the study leverages advanced natural language processing (NLP) techniques to foster a paradigm shift towards more sophisticated data processing systems. Through this comparative study, we aspire to unveil a more streamlined, precise, and efficient approach to document processing in the healthcare domain, potentially revolutionizing the way medical data is handled and analyzed., Comment: 9 pages, 2 figures, 3 tables

As the capabilities of Large Language Models (LLMs) expand, it becomes increasingly important to evaluate them beyond basic knowledge assessment, focusing on higher-level language understanding. This study introduces MultiPragEval, the first multilingual pragmatic evaluation of LLMs, designed for English, German, Korean, and Chinese. Comprising 1200 question units categorized according to Grice's Cooperative Principle and its four conversational maxims, MultiPragEval enables an in-depth assessment of LLMs' contextual awareness and their ability to infer implied meanings. Our findings demonstrate that Claude3-Opus significantly outperforms other models in all tested languages, establishing a state-of-the-art in the field. Among open-source models, Solar-10.7B and Qwen1.5-14B emerge as strong competitors. By analyzing pragmatic inference, we provide valuable insights into the capabilities essential for advanced language comprehension in AI systems., Comment: The 2nd GenBench workshop on generalisation (benchmarking) in NLP

This paper presents initial flight results for distributed optical angles-only navigation of a swarm of small spacecraft, conducted during the Starling Formation-Flying Optical Experiment (StarFOX). StarFOX is a core payload of the NASA Starling mission, which consists of four CubeSats launched in 2023. Prior angles-only flight demonstrations have only featured one observer and target and have relied upon a-priori target orbit knowledge for initialization, translational maneuvers to resolve target range, and external absolute orbit updates to maintain convergence. StarFOX overcomes these limitations by applying the angles-only Absolute and Relative Trajectory Measurement System (ARTMS), which integrates three novel algorithms. Image Processing detects and tracks multiple targets in images from each satellite's on-board camera. Batch Orbit Determination computes initial swarm orbit estimates from bearing angle batches. Sequential Orbit Determination leverages an unscented Kalman filter to refine swarm state estimates over time. Multi-observer measurements shared over an intersatellite link are seamlessly fused to enable absolute and relative orbit determination. StarFOX flight data presents the first demonstrations of autonomous angles-only navigation for a satellite swarm, including multi-target and multi-observer relative navigation; autonomous initialization of navigation for unknown targets; and simultaneous absolute and relative orbit determination. Relative positioning uncertainties of 1.3% of target range (1$\sigma$) are achieved for a single observer under challenging measurement conditions, reduced to 0.6% (1$\sigma$) with multiple observers. Results demonstrate promising performance with regards to ongoing StarFOX campaigns and the application of angles-only navigation to future distributed missions., Comment: Accepted to the 38th Small Satellite Conference

Crystals with unique ionic arrangements and strong electronic correlations serve as a fertile ground for the emergence of exotic phases, as evidenced by the coexistence of charge density wave (CDW) and superconductivity in vanadium Kagome metals, specifically AV3Sb5 (where A represents K, Rb, or Cs). The formation of a star of David CDW superstructure, resulting from the coordinated displacements of vanadium ions on a corner sharing triangular lattice, has garnered significant attention in efforts to comprehend the influence of electron phonon interaction within this geometrically intricate lattice. However, understanding of the underlying mechanism behind CDW formation, coupled with symmetry protected lattice vibrations, remains elusive. In this study, we employed time resolved X ray scattering experiments utilising an X ray free electron laser. Our findings reveal that the phonon mode associated with the out of plane motion of Cs ions becomes frustrated in the CDW phase. Furthermore, we observed the photoinduced emergence of a metastable CDW phase, facilitated by the alleviation of frustration through nonadiabatic changes in free energy. By elucidating the longstanding puzzle surrounding the intervention of phonons in CDW ordering, this research offers fresh insights into the competition between phonons and periodic lattice distortions, a phenomenon widespread in other correlated quantum materials including layered high Tc superconductors., Comment: Manuscript: 20 pages, 4 figures, SI: 14 pages, 8 figures

Vision Transformers (ViT), when paired with large-scale pretraining, have shown remarkable performance across various computer vision tasks, primarily due to their weak inductive bias. However, while such weak inductive bias aids in pretraining scalability, this may hinder the effective adaptation of ViTs for visuo-motor control tasks as a result of the absence of control-centric inductive biases. Such absent inductive biases include spatial locality and translation equivariance bias which convolutions naturally offer. To this end, we introduce Convolution Injector (CoIn), an add-on module that injects convolutions which are rich in locality and equivariance biases into a pretrained ViT for effective adaptation in visuo-motor control. We evaluate CoIn with three distinct types of pretrained ViTs (CLIP, MVP, VC-1) across 12 varied control tasks within three separate domains (Adroit, MetaWorld, DMC), and demonstrate that CoIn consistently enhances control task performance across all experimented environments and models, validating the effectiveness of providing pretrained ViTs with control-centric biases., Comment: accepted to ICML 2024

Recently, various pre-training methods have been introduced in vision-based Reinforcement Learning (RL). However, their generalization ability remains unclear due to evaluations being limited to in-distribution environments and non-unified experimental setups. To address this, we introduce the Atari Pre-training Benchmark (Atari-PB), which pre-trains a ResNet-50 model on 10 million transitions from 50 Atari games and evaluates it across diverse environment distributions. Our experiments show that pre-training objectives focused on learning task-agnostic features (e.g., identifying objects and understanding temporal dynamics) enhance generalization across different environments. In contrast, objectives focused on learning task-specific knowledge (e.g., identifying agents and fitting reward functions) improve performance in environments similar to the pre-training dataset but not in varied ones. We publicize our codes, datasets, and model checkpoints at https://github.com/dojeon-ai/Atari-PB., Comment: accepted to ICML 2024