A naukri.com initiative
ML News
Arxiv
2d
205
Image Credit: Arxiv
SLED: A Speculative LLM Decoding Framework for Efficient Edge Serving
- Efficient inferencing of large language models at the edge is challenging due to device limitations.
- Existing strategies like quantization and pruning trade accuracy for efficiency or incur high costs.
- A new approach called SLED leverages speculative decoding for efficient edge serving.
- SLED orchestrates computation across heterogeneous devices for edge computing.
- The method allows lightweight edge devices to draft multiple candidate tokens locally using diverse models.
- A shared edge server efficiently batches and verifies tokens using a precise model.
- SLED supports device heterogeneity and reduces server-side memory footprint by avoiding multiple target models.
- Initial experiments with Jetson Orin Nano, Raspberry Pi 5, and an RTX 6000 edge server show reduced latency, improved energy efficiency, and increased inference sessions.
- The benefits are achieved without sacrificing model accuracy.
Read Full Article
12 Likes
Arxiv
2d
280
Image Credit: Arxiv
Scoop-and-Toss: Dynamic Object Collection for Quadrupedal Systems
- Quadruped robots have advanced in locomotion and are exploring loco-manipulation using their legs for tasks like pressing buttons.
- A framework is proposed to enable quadruped robots to collect objects by attaching a scoop-like add-on to one leg.
- The robots can scoop objects and toss them into a collection tray on their back without additional actuators.
- The method involves a hierarchical policy structure with expert policies for scooping and tossing, and approaching object positions, along with a meta-policy for dynamic switching.
- Expert policies are trained separately, followed by meta-policy training for coordinated multi-object collection.
- This approach showcases the effective utilization of quadruped legs for dynamic object manipulation.
- Title: Scoop-and-Toss: Dynamic Object Collection for Quadrupedal Systems
- Summary Source: arXiv:2506.09406v1
- Type: cross
- Abstract: Quadruped robots have made significant advances in locomotion, extending their capabilities from controlled environments to real-world applications.
- Recent work explores loco-manipulation using legs for tasks such as pressing buttons or opening doors.
- A framework is proposed to enable quadruped robots to collect objects without additional actuators by using a scoop-like add-on on one leg.
- The method includes a hierarchical policy structure with expert policies for various tasks and a meta-policy for dynamic switching between them.
- This approach demonstrates effective utilization of quadruped legs for dynamic object manipulation beyond locomotion.
- Robots can scoop objects and toss them into a mounted tray using this approach.
- Expert policies and meta-policy training play key roles in coordinated object collection by quadruped robots.
Read Full Article
16 Likes
Arxiv
2d
59
Image Credit: Arxiv
A Call for Collaborative Intelligence: Why Human-Agent Systems Should Precede AI Autonomy
- Recent improvements in large language models (LLMs) have prompted a focus on building fully autonomous AI agents, but a position paper argues against this approach.
- Autonomous AI systems still face challenges with reliability, transparency, and understanding human requirements.
- The paper proposes LLM-based Human-Agent Systems (LLM-HAS) where AI collaborates with humans instead of replacing them, ensuring trustworthiness and adaptability.
- By involving humans to provide guidance and maintain control, these systems can be more reliable and suitable for tasks like healthcare, finance, and software development.
- Human-AI teamwork is shown to handle complex tasks more effectively than AI working in isolation, with examples provided from various industries.
- Challenges in building collaborative systems are discussed in the paper along with practical solutions to address them.
- Progress in AI, as argued in the paper, should be based on how effectively systems can work with humans rather than becoming fully independent.
- The paper advocates for AI systems that enhance human capabilities through partnership rather than replacing human roles entirely.
Read Full Article
3 Likes
Arxiv
2d
51
Image Credit: Arxiv
Time-Unified Diffusion Policy with Action Discrimination for Robotic Manipulation
- A new Time-Unified Diffusion Policy (TUDP) has been developed for robotic manipulation to efficiently generate robot actions with high accuracy.
- The TUDP integrates action recognition capabilities to streamline the action denoising process while enhancing training efficiency and speeding up action generation.
- It introduces a time-unified velocity field in action space with action discrimination information to simplify policy learning and improve action generation speed.
- The TUDP also implements an action-wise training method that includes an action discrimination branch to enhance successful action recognition and denoising accuracy.
- This novel method achieved state-of-the-art performance on RLBench with success rates of 82.6% on a multi-view setup and 83.8% on a single-view setup.
- When using fewer denoising iterations, TUDP demonstrated a significant improvement in success rate, showcasing its efficiency.
- The TUDP is capable of producing accurate actions for various real-world tasks, making it a versatile and reliable solution for robotic manipulation.
Read Full Article
3 Likes
Arxiv
2d
315
Image Credit: Arxiv
When Is Diversity Rewarded in Cooperative Multi-Agent Learning?
- The success of teams in robotics, nature, and society relies on diversified specialists, but a clear explanation of when diversity is better than uniformity is lacking.
- Research focuses on reward design in multi-agent task allocation problems to determine the effectiveness of heterogeneous teams.
- The study examines different types of reward objectives for heterogeneous teams in a non-spatial setting.
- They use generalized aggregation operators to determine whether heterogeneity can enhance rewards.
- It is proven that the curvature of these operators affects how heterogeneity impacts rewards, with a simple convexity test for broad reward families.
- The study investigates how heterogeneity emerges in embodied agents learning effort allocation policies.
- Multi-Agent Reinforcement Learning (MARL) and Heterogeneous Environment Design (HED) are used to optimize scenarios where diversity is beneficial.
- Experiments in matrix games and a Multi-Goal-Capture environment confirm that HED aligns with theoretical predictions regarding the advantages of heterogeneity.
- The findings contribute to understanding when behavioral diversity leads to tangible benefits in cooperative multi-agent learning.
Read Full Article
19 Likes
Arxiv
2d
205
Image Credit: Arxiv
Attention-Bayesian Hybrid Approach to Modular Multiple Particle Tracking
- Tracking multiple particles in noisy and cluttered scenes is challenging due to trajectory hypothesis combinatorial explosion.
- The transformer architecture improves robustness but falls short in scenarios with a reduced set of trajectory hypotheses.
- A hybrid approach combining self-attention of transformers with Bayesian filtering's reliability and interpretability is introduced.
- Trajectory-to-detection association is done by solving a label prediction problem using a transformer encoder.
- This hybrid approach prunes the hypothesis set, enabling efficient multiple-particle tracking in a Bayesian filtering framework.
- The approach shows improved tracking accuracy and robustness against spurious detections in high clutter scenarios.
Read Full Article
12 Likes
Arxiv
2d
59
Image Credit: Arxiv
Towards Bridging the Reward-Generation Gap in Direct Alignment Algorithms
- Direct Alignment Algorithms (DAAs) like Direct Preference Optimization and Simple Preference Optimization have become efficient alternatives to Reinforcement Learning from Human Feedback algorithms for aligning large language models with human preferences.
- DAAs face a limitation known as the 'reward-generation gap,' which represents a misalignment between training optimization objectives and generation performance during inference.
- One contributor to the reward-generation gap is the discrepancy in how prefix tokens' importance affects LLM generation and the implicit reward functions of DAAs.
- To address this gap, a method called Prefix-Oriented Equal-length Training (POET) is introduced, which truncates both preferred and dispreferred responses to match the shorter one's length.
- By training with POET, the optimization of DAAs is constrained to converge across all positions, paying more attention to prefix tokens compared to standard DAAs.
- Experiments with DPO and SimPO, two typical DAAs, show that POET enhances their performance, yielding improvements of up to 15.6 points in AlpacaEval 2 and overall enhancements in downstream tasks.
- The study emphasizes the significance of addressing the gap between reward optimization and generation performance in DAAs.
Read Full Article
3 Likes
Arxiv
2d
86
Image Credit: Arxiv
BemaGANv2: A Tutorial and Comparative Survey of GAN-based Vocoders for Long-Term Audio Generation
- This paper introduces BemaGANv2, an advanced GAN-based vocoder for high-fidelity and long-term audio generation.
- BemaGANv2 builds upon the original BemaGAN architecture by incorporating architectural innovations like the Anti-aliased Multi-Periodicity composition (AMP) module in the generator.
- The generator in BemaGANv2 uses the Snake activation function to better model periodic structures in audio.
- BemaGANv2's discriminator framework includes the Multi-Envelope Discriminator (MED) to extract temporal envelope features and the Multi-Resolution Discriminator (MRD) to model long-range dependencies.
- The evaluation of BemaGANv2 includes different discriminator configurations like MSD + MED, MSD + MRD, and MPD + MED + MRD using various objective metrics and subjective evaluations.
- Objective metrics used for evaluation include FAD, SSIM, PLCC, and MCD, while subjective evaluations involve MOS and SMOS scores.
- The paper provides a tutorial on model architecture, training methodology, and implementation details to ensure reproducibility.
- The code and pre-trained models for BemaGANv2 are available at https://github.com/dinhoitt/BemaGANv2.
Read Full Article
5 Likes
Arxiv
2d
367
Image Credit: Arxiv
Bridging Online Behavior and Clinical Insight: A Longitudinal LLM-based Study of Suicidality on YouTube Reveals Novel Digital Markers
- Suicide remains a leading cause of death in Western countries, prompting new research approaches.
- Digital footprints on social media can provide insight into suicidal behavior.
- A study focused on individuals who attempted suicide while uploading videos to their YouTube channels.
- Complementary approaches were applied: computational bottom-up, hybrid, and expert-driven top-down.
- A novel longitudinal dataset of 181 YouTube channels from individuals with life-threatening attempts was analyzed.
- LLM-based topic modeling identified five behavioral indicators associated with suicide attempts.
- Two topics showed temporal attempt-related changes: Mental Health Struggles and YouTube Engagement.
- A clinical expert reviewed topics and flagged 19 as suicide-related but found no significant attempt-related temporal effects.
- YouTube Engagement, a platform-specific indicator, was not identified by the expert.
- Psychological assessment of suicide attempt narratives revealed differences in motivation to share experiences.
- Individuals who attempted before aimed to Help Others, while those who attempted during their upload period framed it as part of their Personal Recovery.
- The study offers a nuanced understanding of suicidality by integrating digital behavior and clinical insights.
Read Full Article
22 Likes
Arxiv
2d
327
Image Credit: Arxiv
A Survey on the Role of Artificial Intelligence and Machine Learning in 6G-V2X Applications
- 6G networks are expected to revolutionize Connected and Autonomous Vehicles (CAVs) by providing ultra-reliable, low-latency, and high-capacity connectivity through Vehicle-to-Everything (V2X) communication.
- Artificial Intelligence (AI) and Machine Learning (ML) play a key role in optimizing V2X communication by enhancing network management, predictive analytics, security, and cooperative driving.
- AI and ML have excelled in domains like natural language processing and computer vision, contributing significantly to the evolution of 6G-V2X applications.
- This survey delves into recent advancements of AI and ML models in the context of 6G-V2X communication, with a focus on techniques like Deep Learning (DL), Reinforcement Learning (RL), Generative Learning (GL), and Federated Learning (FL).
- Notably, Generative Learning (GL) has shown remarkable progress in enhancing the performance, adaptability, and intelligence of 6G-V2X systems.
- The survey aims to address the lack of a systematic summary of recent research efforts, analyzing the roles of AI and ML in intelligent resource allocation, beamforming, traffic management, and security within 6G-V2X applications.
- Challenges such as computational complexity, data privacy, and real-time decision-making constraints are explored, alongside future research directions to drive AI-driven 6G-V2X development.
- The study provides valuable insights for researchers, engineers, and policymakers involved in the advancement of intelligent, AI-powered V2X ecosystems in 6G communication.
Read Full Article
19 Likes
Arxiv
2d
312
Image Credit: Arxiv
LLM-Powered CPI Prediction Inference with Online Text Time Series
- Forecasting the Consumer Price Index (CPI) is an essential yet complex task in economics, often relying on survey-based data.
- This paper introduces LLM-CPI, an approach that utilizes large language models (LLMs) to improve CPI prediction by incorporating high-frequency online text data.
- LLMs like ChatGPT and BERT are used to generate continuous inflation labels from online texts collected from a Chinese social network site.
- Online text embeddings are obtained through LDA and BERT techniques.
- A joint time series framework is developed that merges monthly CPI data with LLM-generated daily CPI surrogates.
- The monthly model combines observed CPI data, text embeddings, and macroeconomic variables in an ARX structure.
- The daily model uses LLM-generated CPI surrogates and text embeddings in a VARX structure.
- The method's asymptotic properties are analyzed, and two forms of prediction intervals are provided.
- The performance and advantages of LLM-CPI are illustrated through simulation and real data examples.
Read Full Article
18 Likes
Arxiv
2d
331
Image Credit: Arxiv
Tightly-Coupled LiDAR-IMU-Leg Odometry with Online Learned Leg Kinematics Incorporating Foot Tactile Information
- Researchers introduce tightly coupled LiDAR-IMU-leg odometry for challenging conditions like featureless environments and deformable terrains.
- They employ an online learning-based leg kinematics model called the neural leg kinematics model, which incorporates foot tactile information.
- The model captures the nonlinear dynamics between robot feet and the ground, enhancing adaptability to weight load changes and terrain conditions.
- Online training of the model ensures adaptability to different scenarios like delivery or transportation tasks.
- The extit{neural adaptive leg odometry factor} and online uncertainty estimation are used for training the kinematics model and odometry estimation on a unified factor graph.
- Real experiments with a quadruped robot in challenging scenarios such as a sandy beach and a campus with various terrains validate the effectiveness of the proposed method.
- The odometry estimation incorporating the extit{neural leg kinematics model} performs better than existing state-of-the-art methods.
- The researchers offer a project page for further information: https://takuokawara.github.io/RAL2025_project_page/
Read Full Article
19 Likes
Arxiv
2d
110
Image Credit: Arxiv
TooBadRL: Trigger Optimization to Boost Effectiveness of Backdoor Attacks on Deep Reinforcement Learning
- Deep reinforcement learning (DRL) has been successful in various domains, but it is susceptible to backdoor attacks during training.
- Existing backdoor attacks in DRL often use simplistic trigger configurations.
- A new framework called TooBadRL focuses on optimizing DRL backdoor triggers in terms of timing, spatial dimensions, and magnitude.
- TooBadRL introduces an adaptive freezing mechanism for injection timing and a cooperative game approach to select influential state variables.
- Furthermore, TooBadRL utilizes a gradient-based adversarial method to optimize injection magnitude under environment constraints.
- Evaluation on three DRL algorithms and nine tasks shows that TooBadRL increases attack success rates while maintaining normal task performance.
- This research emphasizes the significance of systematic trigger optimization in DRL backdoor attacks.
- The TooBadRL framework's source code is available at https://github.com/S3IC-Lab/TooBadRL.
Read Full Article
6 Likes
Arxiv
2d
102
Image Credit: Arxiv
Learning Efficient and Generalizable Graph Retriever for Knowledge-Graph Question Answering
- Large Language Models (LLMs) suffer from outdated knowledge and hallucinations, hindering their reliability.
- Retrieval-Augmented Generation helps ground LLMs with external knowledge, but current pipelines mostly use unstructured text, limiting interpretability and structured reasoning.
- Knowledge graphs offer a more structured and compact alternative to unstructured text, representing facts as relational triples.
- Recent studies have integrated knowledge graphs with LLMs for Knowledge Graph Question Answering (KGQA), often using a retrieve-then-reason paradigm.
- Graph-based retrievers in KGQA have shown strong empirical performance but struggle with generalization ability.
- A new framework called RAPL is proposed for efficient and effective graph retrieval in KGQA.
- RAPL addresses limitations through a two-stage labeling strategy, a model-agnostic graph transformation approach, and a path-based reasoning strategy.
- The two-stage labeling strategy combines heuristic signals with parametric models to provide causally grounded supervision.
- The model-agnostic graph transformation captures intra- and inter-triple interactions, enhancing representational capacity.
- The path-based reasoning strategy enables learning from rational knowledge injections and supports downstream reasoners with structured inputs.
- Empirically, RAPL outperforms state-of-the-art methods by 2.66%-20.34% and reduces the performance gap between different LLM-based reasoners as well as in cross-dataset settings.
- The framework highlights superior retrieval capability and generalizability in KGQA.
- Codes for RAPL are available at: https://github.com/tianyao-aka/RAPL.
Read Full Article
6 Likes
Arxiv
2d
94
Image Credit: Arxiv
Real-Time Network Traffic Forecasting with Missing Data: A Generative Model Approach
- Real-time network traffic forecasting is essential for network management and resource allocation.
- Existing approaches assume full network traffic data, but practical scenarios often have missing data.
- A generative model approach is proposed for real-time network traffic forecasting with missing data.
- The approach models forecasting as a tensor completion problem and incorporates a pre-trained generative model for low-rank structure.
- The generative model captures the low-rank structure of network traffic data, simplifying the optimization process.
- Optimization is done on the latent representation rather than the high-dimensional tensor.
- A theoretical recovery guarantee quantifies the error bound of the proposed approach.
- Experiments on real-world datasets show accurate network traffic forecasting within 100 ms with a MAE below 0.002.
Read Full Article
5 Likes
For uninterrupted reading, download the app