A naukri.com initiative
ML News
Arxiv
1d
367
Image Credit: Arxiv
De-AntiFake: Rethinking the Protective Perturbations Against Voice Cloning Attacks
- Speech synthesis advancements have raised concerns about voice cloning attacks despite existing protective perturbations.
- While previous studies focused on disrupting voice cloning with perturbations, determined attackers could overcome these defenses.
- A new study evaluates protective perturbations against voice cloning with perturbation purification, showcasing limitations in current methods.
- A novel two-stage purification method, including phoneme guidance, demonstrates improved disruption of voice cloning defenses.
Read Full Article
22 Likes
Arxiv
1d
371
Image Credit: Arxiv
Alleviating Attack Data Scarcity: SCANIA's Experience Towards Enhancing In-Vehicle Cyber Security Measures
- The need for in-vehicle cyber security measures in connected vehicles is critical due to increasing security risks.
- Implementing intrusion detection and response systems with adaptive detection mechanisms is essential to detect evolving threats.
- Constraints on implementing diverse attack scenarios on test vehicles lead to a scarcity of attack-representing data.
- A context-aware attack data generator was developed to efficiently create high-quality attack-representing data for in-vehicle network security testing.
Read Full Article
22 Likes
Arxiv
1d
147
Image Credit: Arxiv
Detection of Disengagement from Voluntary Quizzes: An Explainable Machine Learning Approach in Higher Distance Education
- This paper presents a study on detecting student disengagement in non-mandatory quizzes in distance education.
- The study involved analyzing data from 42 courses over four semesters from a distance-based university using machine learning algorithms.
- An explainable machine learning framework was developed to aid in understanding algorithm decisions, achieving a balanced accuracy of 91% in detecting disengaged students.
- The research also discusses strategies for timely intervention to reduce disengagement in online learning tasks.
Read Full Article
8 Likes
Arxiv
1d
166
Image Credit: Arxiv
Learning few-step posterior samplers by unfolding and distillation of diffusion models
- Diffusion models (DMs) are powerful image priors in Bayesian computational imaging.
- Two primary strategies for leveraging DMs are Plug-and-Play methods and specialized conditional DMs.
- A novel framework integrating deep unfolding and model distillation transforms a DM image prior into a few-step conditional model for posterior sampling.
- The approach includes unfolding a Markov chain Monte Carlo algorithm and achieves excellent accuracy, efficiency, and flexibility in adapting to variations in the forward model.
Read Full Article
10 Likes
Arxiv
1d
263
Image Credit: Arxiv
RLHGNN: Reinforcement Learning-driven Heterogeneous Graph Neural Network for Next Activity Prediction in Business Processes
- Next activity prediction in business processes is crucial for optimizing service-oriented architectures like microservices environments and distributed enterprise systems.
- Existing sequence-based methods struggle to capture non-sequential relationships arising from parallel executions and conditional dependencies, while graph-based approaches lack adaptability due to homogeneous representations and static structures.
- A novel framework called RLHGNN is introduced to transform event logs into heterogeneous process graphs, offering flexible graph structures tailored to individual process complexities through reinforcement learning and heterogeneous graph convolution.
- RLHGNN has demonstrated superior performance over existing methods in predicting next activities, with minimal latency, making it a practical solution for real-time business process monitoring applications.
Read Full Article
15 Likes
Arxiv
1d
69
Image Credit: Arxiv
Bourbaki: Self-Generated and Goal-Conditioned MDPs for Theorem Proving
- Large language models face challenges in automated theorem proving due to sparse rewards and complex reasoning tasks.
- A new framework called self-generated goal-conditioned MDPs (sG-MDPs) is introduced to tackle these challenges by allowing agents to generate and pursue subgoals in a structured manner.
- Monte Carlo Tree Search (MCTS)-like algorithms are utilized to solve the sG-MDP, implemented in Bourbaki (7B) system, which utilizes multiple LLMs for subgoal generation and tactic synthesis.
- Bourbaki (7B) achieves state-of-the-art results on PutnamBench by solving 26 problems, demonstrating the effectiveness of the approach in theorem proving.
Read Full Article
4 Likes
Arxiv
1d
193
Image Credit: Arxiv
Early Signs of Steganographic Capabilities in Frontier LLMs
- Monitoring Large Language Model (LLM) outputs is important to prevent misuse and misalignment.
- LLMs could use steganography to hide information in seemingly innocent text.
- Research found that current LLMs struggle to hide short messages but can do so with specific conditions like an unmonitored scratchpad.
- Despite limited steganographic capabilities, there are early indications that LLMs can perform basic encoded reasoning.
Read Full Article
11 Likes
Arxiv
1d
317
Image Credit: Arxiv
Linear Attention with Global Context: A Multipole Attention Mechanism for Vision and Physics
- Transformers are widely used for tasks like image classification and physics simulations, but their quadratic complexity makes them impractical for high-resolution inputs.
- A new approach called Multipole Attention Neural Operator (MANO) is introduced to address this issue by computing attention in a distance-based multiscale fashion.
- MANO maintains a global receptive field in each attention head, achieving linear time and memory complexity with respect to the number of grid points.
- Empirical results show that MANO competes with state-of-the-art models like ViT and Swin Transformer, reducing runtime and peak memory usage significantly.
Read Full Article
19 Likes
Arxiv
1d
270
Image Credit: Arxiv
KERAP: A Knowledge-Enhanced Reasoning Approach for Accurate Zero-shot Diagnosis Prediction Using Multi-agent LLMs
- Medical diagnosis prediction is crucial for disease detection and personalized healthcare.
- Machine learning models face limitations in generalizing to unseen cases due to their reliance on supervised training and the need for large labeled datasets.
- A new approach called KERAP, a knowledge graph-enhanced reasoning method, addresses challenges faced by large language models in diagnosis prediction by using a multi-agent architecture.
- Experimental results show that KERAP enhances diagnostic reliability and offers a scalable and interpretable solution for zero-shot medical diagnosis prediction.
Read Full Article
16 Likes
Arxiv
1d
290
Image Credit: Arxiv
Self-Correction Bench: Revealing and Addressing the Self-Correction Blind Spot in LLMs
- Large language models (LLMs) can make mistakes and struggle with self-correction, leading to a 'Self-Correction Blind Spot.'
- Researchers introduce the Self-Correction Bench framework to measure the blind spot by injecting controlled errors at varying complexity levels.
- Testing 14 models revealed an average blind spot rate of 64.5%, with training data composition playing a crucial role in this limitation.
- Appending the word 'Wait' reduced blind spots by 89.3%, showing potential for improving the reliability and trustworthiness of LLMs.
Read Full Article
17 Likes
Arxiv
1d
116
Image Credit: Arxiv
Self-Steering Deep Non-Linear Spatially Selective Filters for Efficient Extraction of Moving Speakers under Weak Guidance
- Recent research focuses on deep non-linear spatially selective filters for enhancing moving speakers efficiently.
- Traditional deep filters excel with known-direction stationary speakers, but struggle with moving speakers without extra tracking algorithms.
- A new method combines a simple particle filter with temporal feedback to enhance speech signals and improve tracking accuracy in dynamic scenarios.
- Evaluation shows the proposed self-steering pipeline outperforms other methods, making it a promising solution for real-time speech enhancement.
Read Full Article
6 Likes
Arxiv
1d
46
Image Credit: Arxiv
Learning to Coordinate Bidders in Non-Truthful Auctions
- In non-truthful auctions like first-price and all-pay auctions, independent strategic behaviors of bidders can lead to undesirable outcomes.
- Coordinating bidders by having a mediator recommend correlated bidding strategies can improve auction systems.
- Learning Bayes correlated equilibria in non-truthful auctions requires understanding bidders' private valuations, which may not always be available.
- The sample complexity of learning Bayes correlated equilibria in auctions can be achieved with a polynomial number of samples.
Read Full Article
2 Likes
Arxiv
1d
197
Image Credit: Arxiv
Measurement as Bricolage: Examining How Data Scientists Construct Target Variables for Predictive Modeling Tasks
- Data scientists often face challenges in translating fuzzy concepts into concrete target variables for predictive modeling tasks.
- A study involving interviews with fifteen data scientists in education and healthcare reveals that they construct target variables through a bricolage process.
- The target variable construction process involves iterative negotiation between high-level measurement objectives and practical constraints to satisfy criteria like validity, simplicity, predictability, portability, and resource requirements.
- Data scientists employ adaptive problem (re)formulation strategies, such as swapping target variables or combining multiple outcomes, to meet modeling objectives effectively.
Read Full Article
11 Likes
Arxiv
1d
290
Image Credit: Arxiv
SynapseRoute: An Auto-Route Switching Framework on Dual-State Large Language Model
- A new framework, SynapseRoute, has been proposed to optimize model selection in large language models for cost and performance balancing.
- Approximately 58% of medical questions can be accurately answered using a low-cost, non-thinking mode without employing high-cost reasoning processes.
- SynapseRoute dynamically routes queries to either the thinking or non-thinking mode based on complexity, leading to improved accuracy, cost-efficiency, and user experience.
- Experimental results on medical datasets show that SynapseRoute enhances overall accuracy, reduces inference time by 36.8%, and decreases token consumption by 39.66% compared to using the thinking mode alone.
Read Full Article
17 Likes
Arxiv
1d
297
Image Credit: Arxiv
DNN-Based Precoding in RIS-Aided mmWave MIMO Systems With Practical Phase Shift
- This paper explores precoding design in millimeter wave (mmWave) multiple-input multiple-output (MIMO) systems using reconfigurable intelligent surfaces (RIS) to enhance transmissions.
- Traditional exhaustive search for optimal codewords in the continuous phase shift is computationally intensive, prompting the use of permuted discrete Fourier transform (DFT) vectors and deep neural networks (DNN) to reduce complexity.
- The DNN approach facilitates faster codeword selection, maintaining sub-optimal spectral efficiency in RIS-aided systems even with variations in the distance between end-users and the RIS.
- Simulation results suggest the efficacy of DNN in enhancing the throughput of mmWave MIMO systems with obstructed direct communication paths.
Read Full Article
17 Likes
For uninterrupted reading, download the app