A naukri.com initiative
ML News
Arxiv
6d
156
Image Credit: Arxiv
Bayesian optimization for mixed variables using an adaptive dimension reduction process: applications to aircraft design
- Multidisciplinary design optimization methods aim at adapting numerical optimization techniques to the design of engineering systems involving multiple disciplines.
- A new approach is proposed for mixed variables constrained Bayesian optimization in the context of aircraft design.
- The approach utilizes a reduction process based on the partial least squares method in constructing surrogate models with less hyperparameters.
- The proposed approach shows a significant improvement compared to genetic algorithms in terms of performance.
Read Full Article
9 Likes
Arxiv
6d
301
Image Credit: Arxiv
SeaView: Software Engineering Agent Visual Interface for Enhanced Workflow
- Auto-regressive LLM-based software engineering (SWE) agents have made significant progress on real-world coding challenges.
- Analysis of SWE agent trajectories is difficult due to long interactions with the environment.
- SeaView is a novel tool designed to assist SWE-agent researchers in visualizing and inspecting their experiments.
- SeaView helps researchers compare experimental runs and identify problems related to LLMs or the environment.
Read Full Article
18 Likes
Arxiv
6d
218
Image Credit: Arxiv
Offline Reinforcement Learning using Human-Aligned Reward Labeling for Autonomous Emergency Braking in Occluded Pedestrian Crossing
- This paper presents a novel pipeline for generating human-aligned reward labels in the context of offline reinforcement learning for autonomous emergency braking in occluded pedestrian crossing scenarios.
- The proposed pipeline addresses the challenge of absent reward signals in real-world datasets by generating labels that reflect human judgment and safety considerations.
- An adaptive safety component is incorporated, allowing the autonomous vehicle to prioritize safety over efficiency in potential collision scenarios by analyzing semantic segmentation maps.
- The results demonstrate the effectiveness of the method in producing reliable and human-aligned reward signals, facilitating the training of autonomous driving systems in alignment with human values.
Read Full Article
13 Likes
Arxiv
6d
345
Image Credit: Arxiv
Generating Fine Details of Entity Interactions
- Generating fine details of entity interactions is a long-standing challenge.
- A new dataset, InterActing, has been introduced with 1000 fine-grained prompts covering different interaction scenarios.
- The proposed approach, DetailScribe, leverages LLMs to decompose interactions and uses a VLM for critiquing generated images.
- The results show significantly improved image quality, indicating the potential of enhanced inference strategies.
Read Full Article
20 Likes
Arxiv
6d
82
Image Credit: Arxiv
Steering CLIP's vision transformer with sparse autoencoders
- Sparse autoencoders (SAEs) have helped address the understanding of vision and language processing mechanisms.
- SAEs trained on CLIP's vision transformer reveal distinct sparsity patterns across layers and token types.
- Metrics are introduced to quantify the steerability of SAE features, with 10-15% of neurons and features being steerable.
- Targeted suppression of SAE features improves performance on vision disentanglement tasks and defense against typographic attacks.
Read Full Article
4 Likes
Arxiv
6d
172
Image Credit: Arxiv
Dimension reduction for derivative-informed operator learning: An analysis of approximation errors
- The study focuses on the derivative-informed learning of nonlinear operators between infinite-dimensional separable Hilbert spaces by neural networks.
- The approximation accuracy of the operator's derivatives can significantly impact the performance of the surrogate model for various outer-loop tasks in science and engineering.
- The study analyzes the approximation errors of neural operators in Sobolev norms over infinite-dimensional Gaussian input measures.
- The analysis is validated on numerical experiments with elliptic PDEs, demonstrating the accuracy of bases informed by the map.
Read Full Article
10 Likes
Arxiv
6d
238
Image Credit: Arxiv
Multi-head Ensemble of Smoothed Classifiers for Certified Robustness
- Randomized Smoothing (RS) is a promising technique for certified robustness in deep neural networks.
- A new approach called SmOothed Multi-head Ensemble (SOME) is proposed, which uses multiple augmented heads with a cosine constraint inside a single DNN.
- SOME achieves improved robustness with reduced computational costs compared to traditional ensemble methods of multiple DNNs.
- Extensive experiments and discussions confirm the effectiveness and efficiency of SOME for certifiably-robust RS-based defense.
Read Full Article
14 Likes
Arxiv
6d
111
Image Credit: Arxiv
Block Verification Accelerates Speculative Decoding
- Speculative decoding is an effective method for lossless acceleration of large language models during inference.
- Block Verification is a simple draft verification algorithm that verifies the entire block jointly, providing additional speedup during inference.
- Block verification improves the wall-clock speed by 5%-8% in various tasks and datasets.
- It maintains the strong lossless guarantee and can be used as a default approach in speculative decoding implementations.
Read Full Article
6 Likes
Arxiv
6d
275
Image Credit: Arxiv
Generalization Error Bounds for Learning under Censored Feedback
- The paper discusses the impacts of censored feedback on generalization error bounds in learning algorithms.
- Censored feedback refers to situations where the true label of a data point is only revealed if a favorable decision is made.
- The paper presents an extension of the Dvoretzky-Kiefer-Wolfowitz inequality to quantify the gap between empirical and theoretical data distribution CDFs in non-IID data due to censored feedback.
- The analysis highlights the need for new error bounds that account for censored feedback to accurately capture a model's generalization guarantees.
Read Full Article
16 Likes
Arxiv
6d
238
Image Credit: Arxiv
Analytical Approximation of the ELBO Gradient in the Context of the Clutter Problem
- Researchers propose an analytical approximation for estimating the gradient of the Evidence Lower Bound (ELBO) in variational inference problems.
- The method specifically addresses the clutter problem, where a Bayesian network consists of observations drawn from a mixture of a Gaussian distribution embedded in unrelated clutter.
- The proposed solution utilizes the reparameterization trick and leverages the assumption that the variational distribution is generally more compactly supported than the Gaussian distribution in the likelihood factors, allowing for efficient local approximation.
- The method demonstrates good accuracy, rate of convergence, and linear computational complexity, making it a promising approach in Bayesian inference.
Read Full Article
14 Likes
Arxiv
6d
69
Image Credit: Arxiv
LUCIE: A Lightweight Uncoupled ClImate Emulator with long-term stability and physical consistency for O(1000)-member ensembles
- Researchers have developed a lightweight climate emulator called LUCIE that can accurately simulate long-term climate conditions.
- LUCIE is fully data-driven and can be trained on as little as 2 years of 6-hourly ERA5 data.
- The emulator remains stable and physically consistent for 100 years with 100 ensemble members.
- LUCIE's simulations match the long-term climatology and variability of temperature, wind, precipitation, and humidity observed in ERA5 data.
Read Full Article
4 Likes
Arxiv
6d
193
Image Credit: Arxiv
Adversarial Attacks on Data Attribution
- Data attribution aims to quantify the contribution of individual training data points to the outputs of an AI model.
- A critical question arises regarding the adversarial robustness of data attribution methods.
- Researchers propose two adversarial attack methods, Shadow Attack and Outlier Attack, to manipulate data-attribution-based compensation.
- Empirical results show significant inflation in data-attribution-based compensation using the proposed attack methods.
Read Full Article
11 Likes
Arxiv
6d
345
Image Credit: Arxiv
Solving Dual Sourcing Problems with Supply Mode Dependent Failure Rates
- This paper discusses dual sourcing problems with supply mode dependent failure rates, specifically in relation to spare parts for downtime-critical assets.
- The paper explores how dual sourcing strategies using conventional and additive manufacturing techniques can optimize sourcing by addressing variations in part properties and failure rates.
- The study proposes an iterative heuristic and reinforcement learning techniques, along with an endogenous parameterized learning (EPL) approach, to manage the distinct failure characteristics of parts produced by different methods.
- In a stylized setting, the best policy achieves an average optimality gap of 0.4% while in an energy sector case study, the policies outperform the baseline in 91.1% of instances with average cost savings of up to 22.6%.
Read Full Article
20 Likes
Arxiv
6d
45
Image Credit: Arxiv
A noise-corrected Langevin algorithm and sampling by half-denoising
- A noise-corrected Langevin algorithm has been proposed for sampling from a given pdf in a real space.
- In deep learning, it is often easier to learn the gradient of the log-density of noisy data, which introduces bias in the Langevin algorithm.
- The noise-corrected Langevin algorithm removes the bias due to noisy data, at least regarding first-order terms.
- The algorithm only requires knowledge of the noisy score function for one single noise level.
Read Full Article
2 Likes
Arxiv
6d
267
Image Credit: Arxiv
Addressing Graph Heterogeneity and Heterophily from A Spectral Perspective
- Graph neural networks (GNNs) face challenges of heterogeneity and heterophily.
- Heterogeneity involves diverse node and edge types in a graph.
- Heterophily refers to connected nodes having dissimilar attributes or labels.
- The proposed Heterogeneous Heterophilic Spectral Graph Neural Network (H2SGNN) addresses these challenges using local and global filtering.
Read Full Article
16 Likes
For uninterrupted reading, download the app