A naukri.com initiative
ML News
Arxiv
2d
206
Image Credit: Arxiv
Offline Trajectory Optimization for Offline Reinforcement Learning
- Offline reinforcement learning aims to learn policies without online exploration by utilizing a dynamics model to generate simulation data for policy learning.
- A new approach called offline trajectory optimization (OTTO) is proposed, which focuses on conducting long-horizon simulations and using model uncertainty to evaluate and correct the generated data.
- OTTO utilizes an ensemble of Transformers known as World Transformers to predict environment dynamics and reward functions, generating long-horizon trajectory simulations and correcting low-confidence data through an uncertainty-based World Evaluator.
- Experiments indicate that OTTO can enhance the performance of offline RL algorithms, even in complex environments with sparse rewards, such as AntMaze.
Read Full Article
12 Likes
Arxiv
2d
257
Image Credit: Arxiv
Solving Probabilistic Verification Problems of Neural Networks using Branch and Bound
- Researchers have developed a new algorithm for solving probabilistic verification problems of neural networks.
- The algorithm is based on computing and refining lower and upper bounds on probabilities over the outputs of a neural network.
- By utilizing advanced bound propagation and branch and bound techniques, the algorithm outperforms existing probabilistic verification methods, reducing solution times significantly.
- Empirical evaluations and theoretical analysis demonstrate the soundness and efficiency of the algorithm in various scenarios, even outperforming dedicated algorithms for specific probabilistic verification problems.
Read Full Article
15 Likes
Arxiv
2d
331
Image Credit: Arxiv
No $D_{\text{train}}$: Model-Agnostic Counterfactual Explanations Using Reinforcement Learning
- Machine learning methods have grown significantly, but their practical use in critical domains is hindered by their opacity.
- Counterfactual explanations (CFEs) offer insights into altering decisions made by ML models, yet existing methods often require access to the model's training dataset.
- A novel model-agnostic CFE method, NTD-CFE, based on reinforcement learning, is introduced to generate explanations without needing the training dataset.
- NTD-CFE is designed for static and multivariate time-series datasets, reducing the search space and making CFEs more actionable by requiring fewer and smaller changes.
Read Full Article
19 Likes
Arxiv
2d
66
Image Credit: Arxiv
Curriculum Negative Mining For Temporal Networks
- Researchers introduce Curriculum Negative Mining (CurNM) to address challenges in training Temporal Graph Neural Networks (TGNNs).
- CurNM is a model-aware curriculum learning framework that adapts the difficulty of negative samples by balancing random, historical, and hard negatives.
- The framework includes a dynamically updated negative pool to overcome challenges like positive sparsity and positive shift in temporal networks.
- Experiments on 12 datasets and 3 TGNNs show that CurNM outperforms baseline methods significantly, with thorough ablation studies confirming its usefulness and robustness.
Read Full Article
3 Likes
Arxiv
2d
229
Image Credit: Arxiv
A Bilevel Optimization Framework for Imbalanced Data Classification
- Researchers propose a new undersampling approach to tackle imbalanced data classification issues by avoiding synthetic data pitfalls and under-fitting.
- Their method selects datapoints based on their potential to improve model loss rather than randomly undersampling majority data.
- The approach aims to identify an optimal subset of majority training data by rejecting redundant datapoints, leveraging a bilevel optimization problem.
- Experimental results demonstrate F1 scores up to 10% higher compared to existing state-of-the-art methods.
Read Full Article
13 Likes
Arxiv
2d
62
Image Credit: Arxiv
A Multi-Granularity Supervised Contrastive Framework for Remaining Useful Life Prediction of Aero-engines
- Accurate remaining useful life (RUL) predictions are crucial for safe operation of aero-engines.
- This paper introduces a multi-granularity supervised contrastive (MGSC) framework to address limitations in current RUL prediction methods.
- The MGSC framework aims to align samples with the same RUL label in the feature space, improving prediction accuracy.
- The proposed strategy is implemented on the CMPASS dataset and enhances RUL prediction accuracy using a convolutional long short-term memory network as a baseline.
Read Full Article
3 Likes
Arxiv
2d
190
Image Credit: Arxiv
Semantic Edge Computing and Semantic Communications in 6G Networks: A Unifying Survey and Research Challenges
- Semantic Edge Computing (SEC) and Semantic Communications (SemComs) are being explored for real-time intelligence in 6G wireless networks.
- SemCom utilizes Deep Neural Networks (DNNs) to encode and communicate semantic information while SEC uses distributed DNNs to optimize computing across devices.
- This work aims to bridge the gap between SEC and SemCom by analyzing research problems, technical strengths, and challenges in both fields.
- The study provides a comprehensive overview of the current state of the art in SEC and SemCom for 6G networks.
Read Full Article
11 Likes
Arxiv
2d
257
Image Credit: Arxiv
Contextual Bandits in Payment Processing: Non-uniform Exploration and Supervised Learning
- Recent research explores the combination of non-uniform exploration and supervised learning in decision-making systems to improve immediate performance while maintaining off-policy learning capabilities.
- An analysis conducted at Adyen, a global payments processor, demonstrates that regression oracles can enhance system performance but may introduce challenges due to rigid algorithmic assumptions.
- The study reveals that improvements in policy may lead to subsequent performance degradation due to shifts in reward distribution and increased class imbalance in training data.
- There is a potential 'oscillation effect' identified where regression oracles influence probability estimates, impacting the stability and performance consistency of policy models over successive iterations.
Read Full Article
15 Likes
Arxiv
2d
264
Image Credit: Arxiv
Challenges learning from imbalanced data using tree-based models: Prevalence estimates systematically depend on hyperparameters and can be upwardly biased
- Imbalanced binary classification problems are common in various fields of study.
- Subsampling the majority class to create a balanced training dataset can bias the model's predictions.
- Calibrating a random forest model using prevalence estimates can lead to unintended negative consequences, including upwardly biased estimates.
- Random forests' prevalence estimates depend on the number of predictors considered at each split and the sampling rate used, revealing unexpected biases.
Read Full Article
15 Likes
Arxiv
2d
381
Image Credit: Arxiv
"I am bad": Interpreting Stealthy, Universal and Robust Audio Jailbreaks in Audio-Language Models
- The paper discusses challenges in machine learning safety introduced by multimodal large language models, with a focus on Audio-Language Models (ALMs).
- It explores audio jailbreaks targeting ALMs, showing the first universal jailbreaks in the audio modality that can bypass alignment mechanisms and remain effective in simulated real-world conditions.
- The research reveals that adversarial perturbations encode imperceptible toxic speech, suggesting that embedding linguistic features within audio signals can elicit toxic outputs.
- The study highlights the importance of understanding interactions between modalities in multimodal models and provides insights to improve defenses against adversarial audio attacks.
Read Full Article
22 Likes
Arxiv
2d
272
Image Credit: Arxiv
Parametric Scaling Law of Tuning Bias in Conformal Prediction
- Conformal prediction is a framework for uncertainty quantification that constructs prediction sets with coverage guarantees, often requiring a holdout set for parameter tuning.
- Empirical findings suggest that the tuning bias, resulting from using the same dataset for tuning and calibration, is minimal for simple parameter tuning in many conformal prediction methods.
- A scaling law for tuning bias is observed, showing that bias increases with parameter space complexity but decreases with calibration set size.
- The study establishes a theoretical framework to quantify tuning bias, provides a proof for the scaling law, and discusses strategies to mitigate tuning bias based on the research findings.
Read Full Article
16 Likes
Arxiv
2d
74
Image Credit: Arxiv
ARBoids: Adaptive Residual Reinforcement Learning With Boids Model for Cooperative Multi-USV Target Defense
- Researchers introduced ARBoids, an adaptive residual reinforcement learning framework for the target defense problem with unmanned surface vehicles (USVs).
- ARBoids integrates deep reinforcement learning (DRL) with the Boids model for multi-agent coordination in challenging interception scenarios.
- In simulations, ARBoids demonstrated superior performance compared to traditional interception strategies and showed adaptability to attackers with varying maneuverability.
- The code for ARBoids will be made available upon the acceptance of this research letter.
Read Full Article
4 Likes
Arxiv
2d
319
Image Credit: Arxiv
Robust and Efficient Writer-Independent IMU-Based Handwriting Recognition
- Online handwriting recognition (HWR) using data from inertial measurement units (IMUs) poses challenges due to writing style variations and limited annotated datasets.
- This paper introduces an HWR model focused on improving writer-independent (WI) recognition on IMU data, employing a CNN encoder and a BiLSTM-based decoder.
- The model exhibits robustness to unseen handwriting styles, surpassing existing methods on WI splits of public datasets with low character error rates (CERs) and word error rates (WERs).
- Extensive evaluation demonstrates its adaptability to different age groups and efficiency through design choices, hinting at the potential for more adaptable and scalable HWR systems.
Read Full Article
19 Likes
Arxiv
2d
350
Image Credit: Arxiv
Deep Learning is Not So Mysterious or Different
- Deep neural networks are often viewed as different due to their generalization behavior, including benign overfitting and double descent.
- The anomalies observed in neural networks can be understood using traditional generalization frameworks like PAC-Bayes.
- The concept of soft inductive biases is key in explaining neural networks' generalization behavior, advocating for a flexible hypothesis space.
- While deep learning shares commonalities with other model classes, it stands out in representation learning and universality.
Read Full Article
21 Likes
Arxiv
2d
101
Image Credit: Arxiv
Mixture of Group Experts for Learning Invariant Representations
- A new perspective on Mixture-of-Experts (MoE) models with top-k routing has been introduced, called Mixture of Group Experts (MoGE), to address limitations of vanilla MoE models.
- MoGE utilizes group sparse regularization for routing inputs, creating a 2D topographic map that enhances expert diversity and specialization, leading to improved performance in tasks like image classification and language modeling.
- Comprehensive evaluations show that MoGE outperforms traditional MoE models with minimal extra memory and computation requirements, offering an efficient solution to scale the number of experts while avoiding redundancy.
- The source code for MoGE is included in the supplementary material and will be made publicly available for further exploration and implementation.
Read Full Article
6 Likes
For uninterrupted reading, download the app