Pruning is widely used to reduce the complexity of deep learning models, but its effects on interpretability and representation learning remain poorly understood. This paper investigates how pruning influences vision models across three key dimensions: (i) interpretability, (ii) unsupervised object discovery, and (iii) alignment with human perception. We first analyze different vision network architectures to examine how varying sparsity levels affect feature attribution interpretability methods. Additionally, we explore whether pruning promotes more succinct and structured representations, potentially improving unsupervised object discovery by discarding redundant information while preserving essential features. Finally, we assess whether pruning enhances the alignment between model representations and human perception, investigating whether sparser models focus on more discriminative features similarly to humans. Our findings also reveal the presence of sweet spots, where sparse models exhibit higher interpretability, downstream generalization and human alignment. However, these spots highly depend on the network architectures and their size in terms of trainable parameters. Our results suggest a complex interplay between these three dimensions, highlighting the importance of investigating when and how pruning benefits vision representations.
In recent years, compressed domain semantic inference has primarily relied on learned image coding models optimized for mean squared error (MSE). However, MSE-oriented optimization tends to yield latent spaces with limited semantic richness, which hinders effective semantic inference in downstream tasks. Moreover, achieving high performance with these models often requires fine-tuning the entire vision model, which is computationally intensive, especially for large models. To address these problems, we introduce Perception-Oriented Latent Coding (POLC), an approach that enriches the semantic content of latent features for high-performance compressed domain semantic inference. With the semantically rich latent space, POLC requires only a plug-and-play adapter for fine-tuning, significantly reducing the parameter count compared to previous MSE-oriented methods. Experimental results demonstrate that POLC achieves rate-perception performance comparable to state-of-the-art generative image coding methods while markedly enhancing performance in vision tasks, with minimal fine-tuning overhead. Code is available at https://github.com/NJUVISION/POLC.
Global warming, loss of biodiversity, and air pollution are among the most significant problems facing Earth. One of the primary challenges in addressing these issues is the lack of monitoring forests to protect them. To tackle this problem, it is important to leverage remote sensing and computer vision methods to automate monitoring applications. Hence, automatic tree crown detection algorithms emerged based on traditional and deep learning methods. In this study, we first introduce two different tree crown detection methods based on these approaches. Then, we form a novel rule-based approach that integrates these two methods to enhance robustness and accuracy of tree crown detection results. While traditional methods are employed for feature extraction and segmentation of forested areas, deep learning methods are used to detect tree crowns in our method. With the proposed rule-based approach, we post-process these results, aiming to increase the number of detected tree crowns through neighboring trees and localized operations. We compare the obtained results with the proposed method in terms of the number of detected tree crowns and report the advantages, disadvantages, and areas for improvement of the obtained outcomes.
With the advent of new sequence models like Mamba and xLSTM, several studies have shown that these models match or outperform state-of-the-art models in single-channel speech enhancement, automatic speech recognition, and self-supervised audio representation learning. However, prior research has demonstrated that sequence models like LSTM and Mamba tend to overfit to the training set. To address this issue, previous works have shown that adding self-attention to LSTMs substantially improves generalization performance for single-channel speech enhancement. Nevertheless, neither the concept of hybrid Mamba and time-frequency attention models nor their generalization performance have been explored for speech enhancement. In this paper, we propose a novel hybrid architecture, MambAttention, which combines Mamba and shared time- and frequency-multi-head attention modules for generalizable single-channel speech enhancement. To train our model, we introduce VoiceBank+Demand Extended (VB-DemandEx), a dataset inspired by VoiceBank+Demand but with more challenging noise types and lower signal-to-noise ratios. Trained on VB-DemandEx, our proposed MambAttention model significantly outperforms existing state-of-the-art LSTM-, xLSTM-, Mamba-, and Conformer-based systems of similar complexity across all reported metrics on two out-of-domain datasets: DNS 2020 and EARS-WHAM_v2, while matching their performance on the in-domain dataset VB-DemandEx. Ablation studies highlight the role of weight sharing between the time- and frequency-multi-head attention modules for generalization performance. Finally, we explore integrating the shared time- and frequency-multi-head attention modules with LSTM and xLSTM, which yields a notable performance improvement on the out-of-domain datasets. However, our MambAttention model remains superior on both out-of-domain datasets across all reported evaluation metrics.
Accurate particle size distribution (PSD) measurement is important in industries such as mining, pharmaceuticals, and fertilizer manufacturing, significantly influencing product quality and operational efficiency. Traditional PSD methods like sieve analysis and laser diffraction are manual, time-consuming, and limited by particle overlap. Recent developments in convolutional neural networks (CNNs) enable automated, real-time PSD estimation directly from particle images. In this work, we present a CNN-based methodology trained on realistic synthetic particle imagery generated using Blender's advanced rendering capabilities. Synthetic data sets using this method can replicate various industrial scenarios by systematically varying particle shapes, textures, lighting, and spatial arrangements that closely resemble the actual configurations. We evaluated three CNN-based architectures, ResNet-50, InceptionV3, and EfficientNet-B0, for predicting critical PSD parameters (d10, d50, d90). Results demonstrated comparable accuracy across models, with EfficientNet-B0 achieving the best computational efficiency suitable for real-time industrial deployment. This approach shows the effectiveness of realistic synthetic data for robust CNN training, which offers significant potential for automated industrial PSD monitoring. The code is released at : https://github.com/YasserElj/Synthetic-Granular-Gen
Zero-shot skeleton-based action recognition aims to classify unseen skeleton-based human actions without prior exposure to such categories during training. This task is extremely challenging due to the difficulty in generalizing from known to unknown actions. Previous studies typically use two-stage training: pre-training skeleton encoders on seen action categories using cross-entropy loss and then aligning pre-extracted skeleton and text features, enabling knowledge transfer to unseen classes through skeleton-text alignment and language models' generalization. However, their efficacy is hindered by 1) insufficient discrimination for skeleton features, as the fixed skeleton encoder fails to capture necessary alignment information for effective skeleton-text alignment; 2) the neglect of alignment bias between skeleton and unseen text features during testing. To this end, we propose a prototype-guided feature alignment paradigm for zero-shot skeleton-based action recognition, termed PGFA. Specifically, we develop an end-to-end cross-modal contrastive training framework to improve skeleton-text alignment, ensuring sufficient discrimination for skeleton features. Additionally, we introduce a prototype-guided text feature alignment strategy to mitigate the adverse impact of the distribution discrepancy during testing. We provide a theoretical analysis to support our prototype-guided text feature alignment strategy and empirically evaluate our overall PGFA on three well-known datasets. Compared with the top competitor SMIE method, our PGFA achieves absolute accuracy improvements of 22.96%, 12.53%, and 18.54% on the NTU-60, NTU-120, and PKU-MMD datasets, respectively.
The architecture of multimodal large language models (MLLMs) commonly connects a vision encoder, often based on CLIP-ViT, to a large language model. While CLIP-ViT works well for capturing global image features, it struggles to model local relationships between adjacent patches, leading to weaker visual representation, which in turn affects the detailed understanding ability of MLLMs. To solve this, we propose LLaVA-SP, which \textbf{ only adds six spatial visual tokens} to the original visual tokens to enhance the visual representation. Our approach offers three key advantages: 1)We propose a novel Projector, which uses convolutional kernels to derive visual spatial tokens from ViT patch features, simulating two visual spatial ordering approaches: ``from central region to global" and ``from abstract to specific". Then, a cross-attention mechanism is applied to fuse fine-grained visual information, enriching the overall visual representation. 2) We present two model variants: LLaVA-SP-Cropping, which focuses on detail features through progressive cropping, and LLaVA-SP-Pooling, which captures global semantics through adaptive pooling, enabling the model to handle diverse visual understanding tasks. 3) Extensive experiments show that LLaVA-SP, fine-tuned with LoRA, achieves significant performance improvements across various multimodal benchmarks, outperforming the state-of-the-art LLaVA-1.5 model in multiple tasks with nearly identical inference latency. The code and models are available at \href{https://github.com/CnFaker/LLaVA-SP}{\texttt{https://github.com/CnFaker/LLaVA-SP}}.
Accurate interpretation of multi-view radiographs is crucial for diagnosing fractures, muscular injuries, and other anomalies. While significant advances have been made in AI-based analysis of single images, current methods often struggle to establish robust correspondences between different X-ray views, an essential capability for precise clinical evaluations. In this work, we present a novel self-supervised pipeline that eliminates the need for manual annotation by automatically generating a many-to-many correspondence matrix between synthetic X-ray views. This is achieved using digitally reconstructed radiographs (DRR), which are automatically derived from unannotated CT volumes. Our approach incorporates a transformer-based training phase to accurately predict correspondences across two or more X-ray views. Furthermore, we demonstrate that learning correspondences among synthetic X-ray views can be leveraged as a pretraining strategy to enhance automatic multi-view fracture detection on real data. Extensive evaluations on both synthetic and real X-ray datasets show that incorporating correspondences improves performance in multi-view fracture classification.
Human Activity Recognition (HAR), which uses data from Inertial Measurement Unit (IMU) sensors, has many practical applications in healthcare and assisted living environments. However, its use in real-world scenarios has been limited by the lack of comprehensive IMU-based HAR datasets that cover a wide range of activities and the lack of transparency in existing HAR models. Zero-shot HAR (ZS-HAR) overcomes the data limitations, but current models struggle to explain their decisions, making them less transparent. This paper introduces a novel IMU-based ZS-HAR model called the Self-Explainable Zero-shot Human Activity Recognition Network (SEZ-HARN). It can recognize activities not encountered during training and provide skeleton videos to explain its decision-making process. We evaluate the effectiveness of the proposed SEZ-HARN on four benchmark datasets PAMAP2, DaLiAc, HTD-MHAD and MHealth and compare its performance against three state-of-the-art black-box ZS-HAR models. The experiment results demonstrate that SEZ-HARN produces realistic and understandable explanations while achieving competitive Zero-shot recognition accuracy. SEZ-HARN achieves a Zero-shot prediction accuracy within 3\% of the best-performing black-box model on PAMAP2 while maintaining comparable performance on the other three datasets.
Deep Reinforcement Learning (DRL) has achieved remarkable success in complex sequential decision-making tasks, such as playing Atari 2600 games and mastering board games. A critical yet underexplored aspect of DRL is the temporal scale of action execution. We propose a novel paradigm that integrates contextual bandits with DRL to adaptively select action durations, enhancing policy flexibility and computational efficiency. Our approach augments a Deep Q-Network (DQN) with a contextual bandit module that learns to choose optimal action repetition rates based on state contexts. Experiments on Atari 2600 games demonstrate significant performance improvements over static duration baselines, highlighting the efficacy of adaptive temporal abstractions in DRL. This paradigm offers a scalable solution for real-time applications like gaming and robotics, where dynamic action durations are critical.
We find an efficient approach to approximately convert matrix product states (MPSs) into restricted Boltzmann machine wave functions consisting of a multinomial hidden unit through a canonical polyadic (CP) decomposition of the MPSs. This method allows us to generate well-behaved initial neural network quantum states for quantum many-body ground-state calculations in polynomial time of the number of variational parameters and systematically shorten the distance between the initial states and the ground states with increasing the rank of the CP decomposition. We demonstrate the efficiency of our method by taking the transverse-field Ising model as an example and discuss possible applications of our method to more general quantum many-body systems in which the ground-state wave functions possess complex nodal structures.
In contrast to grid-following inverters, Virtual Synchronous Generators (VSGs) perform well under weak grid conditions but may become unstable when the grid is strong. Grid strength depends on grid impedance, which unfortunately varies over time. In this paper, we propose a novel adaptive gain-scheduling control scheme for VSGs. First, an Artificial Neural Network (ANN) estimates the fundamental-frequency grid impedance; then these estimates are fed into an adaptive gain-scheduling function to recalculate controller parameters under varying grid conditions. The proposed method is validated in Simulink and compared with a conventional VSG employing fixed controller gains. The results demonstrate that settling times and overshoot percentages remain consistent across different grid conditions. Additionally, previously unseen grid impedance values are estimated with high accuracy and minimal time delay, making the approach well suited for real-time gain-scheduling control.
Federated learning(FL) is used for distributed scenarios to train artificial intelligence(AI) models while ensuring users' privacy. In federated learning scenario, the server generally never knows about users' data. This type of concept makes the AI training process efficient in terms of data privacy. However, regarding model performance, federated AI models may not sufficiently satisfy AI users' expectations. Furthermore, AI users have a wide range of different needs. It is not easy to satisfy the whole users needs. These types of issues can be addressed through AI model optimization, fine-tuning, or personalization to achieve optimal model performance. To address model optimization challenges, we propose reference model-based federated learning for optimal fine-tuning, which overcomes catastrophic forgetting in each round. This method is derived from Bayesian parameter-efficient transfer learning, which includes an optimal proximal term and enables overcoming the catastrophic forgetting issue in each round by utilizing a reference model that incorporates previous model parameters. As a result, this method achieves both high model performance and low computing cost.
We present Seg-R1, a preliminary exploration of using reinforcement learning (RL) to enhance the pixel-level understanding and reasoning capabilities of large multimodal models (LMMs). Starting with foreground segmentation tasks, specifically camouflaged object detection (COD) and salient object detection (SOD), our approach enables the LMM to generate point and bounding box prompts in the next-token fashion, which are then used to guide SAM2 in producing segmentation masks. We introduce Group Relative Policy Optimization (GRPO) into the segmentation domain, equipping the LMM with pixel-level comprehension through a carefully designed training strategy. Notably, Seg-R1 achieves remarkable performance with purely RL-based training, achieving .873 S-measure on COD10K without complex model modification. Moreover, we found that pure RL training demonstrates strong open-world generalization. Despite being trained solely on foreground segmentation image-mask pairs without text supervision, Seg-R1 achieves impressive zero-shot performance on referring segmentation and reasoning segmentation tasks, with 71.4 cIoU on RefCOCOg test and 56.7 gIoU on ReasonSeg test, outperforming models fully supervised on these datasets.
Hitchhiking, a spontaneous and decentralized mode of travel, has long eluded systematic study due to its informal nature. This paper presents and analyzes the largest known structured dataset of hitchhiking rides, comprising over 63,000 entries collected over nearly two decades through platforms associated with hitchwiki.org and lately on hitchmap.com. By leveraging crowd-sourced contributions, the dataset captures key spatiotemporal and strategic aspects of hitchhiking. This work documents the dataset's origins, evolution, and community-driven maintenance, highlighting its Europe-centric distribution, seasonal patterns, and reliance on a small number of highly active contributors. Through exploratory analyses, I examine waiting times, user behavior, and comment metadata, shedding light on the lived realities of hitchhikers. While the dataset has inherent biases and limitations - such as demographic skew and unverifiable entries it offers a rare and valuable window into an alternative form of mobility. I conclude by outlining future directions for enriching the dataset and advancing research on hitchhiking as both a transportation practice and cultural phenomenon.
In this paper, we present LLaVA-Scissor, a training-free token compression strategy designed for video multimodal large language models. Previous methods mostly attempt to compress tokens based on attention scores, but fail to effectively capture all semantic regions and often lead to token redundancy. Differently, we propose to leverage the Semantic Connected Components (SCC) approach that assigns tokens to distinct semantic regions within the token set, ensuring comprehensive semantic coverage. The outcome is a two-step spatio-temporal token compression strategy that utilizes SCC in both spatial and temporal domains. This strategy can effectively compress tokens by representing the entire video with a set of non-overlapping semantic tokens. We conduct extensive evaluations of the token compression capabilities of LLaVA-Scissor across diverse video understanding benchmarks, including video question answering, long video understanding, and comprehensive multi-choices benchmarks. Experimental results show that the proposed LLaVA-Scissor outperforms other token compression methods, achieving superior performance in various video understanding benchmarks, particularly at low token retention ratios. Project page: https://github.com/HumanMLLM/LLaVA-Scissor.
Simulating showers of particles in highly-granular calorimeters is a key frontier in the application of machine learning to particle physics. Achieving high accuracy and speed with generative machine learning models can enable them to augment traditional simulations and alleviate a major computing constraint. Recent developments have shown how diffusion based generative shower simulation approaches that do not rely on a fixed structure, but instead generate geometry-independent point clouds, are very efficient. We present a transformer-based extension to previous architectures which were developed for simulating electromagnetic showers in the highly granular electromagnetic calorimeter of the International Large Detector, ILD. The attention mechanism now allows us to generate complex hadronic showers with more pronounced substructure across both the electromagnetic and hadronic calorimeters. This is the first time that machine learning methods are used to holistically generate showers across the electromagnetic and hadronic calorimeter in highly granular imaging calorimeter systems.
Text-to-SQL enables non-experts to retrieve data from databases by converting natural language queries into SQL. However, state-of-the-art text-to-SQL studies rely on the BIRD dataset, which assumes that evidence is provided along with questions. Although BIRD facilitates research advancements, it assumes that users have expertise and domain knowledge, contradicting the fundamental goal of text-to-SQL. In addition, human-generated evidence in BIRD contains defects, including missing or erroneous evidence, which affects model performance. To address this issue, we propose SEED (System for Evidence Extraction and Domain knowledge generation), an approach that automatically generates evidence to improve performance and practical usability in real-world scenarios. SEED systematically analyzes database schema, description files, and values to extract relevant information. We evaluated SEED on BIRD and Spider, demonstrating that it significantly improves SQL generation accuracy in the no-evidence scenario, and in some cases, even outperforms the setting where BIRD evidence is provided. Our results highlight that SEED-generated evidence not only bridges the gap between research and real-world deployment but also improves the adaptability and robustness of text-to-SQL models. Our code is available at https://github.com/felix01189/SEED
Skill-based reinforcement learning (SBRL) enables rapid adaptation in environments with sparse rewards by pretraining a skill-conditioned policy. Effective skill learning requires jointly maximizing both exploration and skill diversity. However, existing methods often face challenges in simultaneously optimizing for these two conflicting objectives. In this work, we propose a new method, Adaptive Multi-objective Projection for balancing Exploration and skill Diversification (AMPED), which explicitly addresses both exploration and skill diversification. We begin by conducting extensive ablation studies to identify and define a set of objectives that effectively capture the aspects of exploration and skill diversity, respectively. During the skill pretraining phase, AMPED introduces a gradient surgery technique to balance the objectives of exploration and skill diversity, mitigating conflicts and reducing reliance on heuristic tuning. In the subsequent fine-tuning phase, AMPED incorporates a skill selector module that dynamically selects suitable skills for downstream tasks, based on task-specific performance signals. Our approach achieves performance that surpasses SBRL baselines across various benchmarks. These results highlight the importance of explicitly harmonizing exploration and diversity and demonstrate the effectiveness of AMPED in enabling robust and generalizable skill learning. Project Page: https://geonwoo.me/amped/
Counterfactual explainability seeks to uncover model decisions by identifying minimal changes to the input that alter the predicted outcome. This task becomes particularly challenging for graph data due to preserving structural integrity and semantic meaning. Unlike prior approaches that rely on forward perturbation mechanisms, we introduce Graph Inverse Style Transfer (GIST), the first framework to re-imagine graph counterfactual generation as a backtracking process, leveraging spectral style transfer. By aligning the global structure with the original input spectrum and preserving local content faithfulness, GIST produces valid counterfactuals as interpolations between the input style and counterfactual content. Tested on 8 binary and multi-class graph classification benchmarks, GIST achieves a remarkable +7.6% improvement in the validity of produced counterfactuals and significant gains (+45.5%) in faithfully explaining the true class distribution. Additionally, GIST's backtracking mechanism effectively mitigates overshooting the underlying predictor's decision boundary, minimizing the spectral differences between the input and the counterfactuals. These results challenge traditional forward perturbation methods, offering a novel perspective that advances graph explainability.