A naukri.com initiative
ML News
Arxiv
1d
168
Image Credit: Arxiv
Breaking Data Silos: Towards Open and Scalable Mobility Foundation Models via Generative Continual Learning
- MoveGCL is a privacy-preserving framework for training mobility foundation models via generative continual learning without sharing raw data.
- It enables decentralized and progressive model evolution by replaying synthetic trajectories generated from a frozen teacher model, while reinforcing knowledge retention through a tailored distillation strategy.
- MoveGCL incorporates a Mixture-of-Experts Transformer with a mobility-aware expert routing mechanism and utilizes a layer-wise progressive adaptation strategy to stabilize continual updates.
- Experiments on six real-world urban datasets show that MoveGCL achieves performance comparable to joint training, surpassing federated learning baselines, and providing strong privacy protection.
Read Full Article
10 Likes
Arxiv
1d
247
Image Credit: Arxiv
Do Protein Transformers Have Biological Intelligence?
- A new study explores the potential of Protein Transformers in capturing biological intelligence among protein sequences.
- The study introduces a protein function dataset named Protein-FN containing over 9000 protein data with meaningful labels.
- A new Transformer architecture called Sequence Protein Transformers (SPT) is developed for efficient protein function predictions.
- The research also introduces an Explainable Artificial Intelligence (XAI) technique called Sequence Score to interpret the decision-making processes of protein models, demonstrating promising results.
Read Full Article
14 Likes
Arxiv
1d
116
Image Credit: Arxiv
A Framework for Controllable Multi-objective Learning with Annealed Stein Variational Hypernetworks
- Pareto Set Learning (PSL) is efficient in Multi-objective Learning (MOL) to obtain the complete optimal solution.
- Approach addresses the challenge of making diverse Pareto solutions while maximizing hypervolume value.
- Proposed method SVH-MOL uses Stein Variational Gradient Descent (SVGD) to approximate entire Pareto set.
- Method validated through experiments on multi-objective problems and multi-task learning, showing superior performance.
Read Full Article
7 Likes
Arxiv
1d
300
Image Credit: Arxiv
The OCR Quest for Generalization: Learning to recognize low-resource alphabets with model editing
- Researchers have developed a method to enhance recognition systems for low-resource alphabets using model editing.
- The aim is to create models that can generalize to new data distributions like alphabets more quickly than current fine-tune strategies.
- The approach leverages model editing advancements to improve low-resource learning by incorporating unseen scripts.
- Experiments show significant performance improvements in transfer learning to new alphabets and out-of-domain evaluation for historical ciphered texts and non-Latin scripts.
Read Full Article
18 Likes
Arxiv
1d
63
Image Credit: Arxiv
Feature-Based Instance Neighbor Discovery: Advanced Stable Test-Time Adaptation in Dynamic World
- Deep neural networks often face performance declines due to distribution shifts between training and test domains, impacting Quality of Experience (QoE) for applications.
- Existing test-time adaptation methods struggle with dynamic, multiple test distributions within batches, revealing limitations in global normalization strategies.
- Feature-based Instance Neighbor Discovery (FIND) is introduced, consisting of Layer-wise Feature Disentanglement (LFD), Feature Aware Batch Normalization (FABN), and Selective FABN (S-FABN) to address these challenges.
- FIND shows significant performance improvements over existing methods, achieving a 30% accuracy enhancement in dynamic scenarios while ensuring computational efficiency.
Read Full Article
3 Likes
Arxiv
1d
45
Image Credit: Arxiv
FuncGNN: Learning Functional Semantics of Logic Circuits with Graph Neural Networks
- FuncGNN is a proposed method to improve the representation of logic circuits using Graph Neural Networks (GNNs).
- It addresses issues such as structural heterogeneity and global logic information loss in And-Inverter Graphs (AIGs) commonly used in electronic design automation.
- FuncGNN integrates hybrid feature aggregation, gate-aware normalization, and multi-layer integration to enhance logic circuit representations.
- Experimental results show that FuncGNN outperforms existing methods in logic-level analysis tasks while reducing training time and GPU memory usage.
Read Full Article
2 Likes
Arxiv
1d
319
Image Credit: Arxiv
Is Optimal Transport Necessary for Inverse Reinforcement Learning?
- Inverse Reinforcement Learning (IRL) aims to retrieve reward function from expert demonstrations.
- Recently, Optimal Transport (OT) methods have been effective but come with complexities.
- A new study challenges the necessity of OT in IRL with two simpler alternatives: Minimum-Distance Reward and Segment-Matching Reward.
- Extensive evaluations show these simple methods match or surpass recent OT-based approaches, indicating a reevaluation of complexity in future IRL design.
Read Full Article
19 Likes
Arxiv
1d
364
Image Credit: Arxiv
IMPA-HGAE:Intra-Meta-Path Augmented Heterogeneous Graph Autoencoder
- IMP-HGAE is a novel framework designed to improve target node embeddings by leveraging internal node information along meta-paths in heterogeneous graphs.
- This approach addresses the limitation of existing models that only utilize information from nodes at the ends of meta-paths.
- IMPA-HGAE has shown superior performance on heterogeneous datasets and introduces masking strategies to enhance generative SSL models on heterogeneous graph data.
- The paper also discusses interpretability of the method and potential future directions for generative self-supervised learning in heterogeneous graphs.
Read Full Article
21 Likes
Arxiv
1d
180
Image Credit: Arxiv
Curvature Enhanced Data Augmentation for Regression
- Deep learning models with a large number of parameters have achieved exceptional performance, thanks to effective regularization techniques like data augmentation.
- While data augmentation is widely used in classification tasks, its application in regression problems has been less explored.
- A novel approach called Curvature-Enhanced Manifold Sampling (CEMS) is introduced for regression tasks, which leverages second-order data manifold representation for efficient sampling.
- CEMS shows superior performance in various datasets and scenarios, with minimal computational overhead, as demonstrated through evaluations and comparisons with existing methods.
Read Full Article
10 Likes
Arxiv
1d
162
Image Credit: Arxiv
High-Fidelity Scientific Simulation Surrogates via Adaptive Implicit Neural Representations
- Implicit neural representations (INRs) are crucial for speeding up scientific simulations but struggle with complex fields with high-frequency variations.
- Feature-Adaptive INR (FA-INR) is proposed as an effective alternative using cross-attention and memory banks for flexible feature representations.
- FA-INR achieves high fidelity in large-scale simulation datasets while reducing model size, striking a new balance between accuracy and compactness.
- Introducing a coordinate-guided mixture of experts (MoE) enhances feature representation specialization and efficiency in FA-INR.
Read Full Article
9 Likes
Arxiv
1d
72
Image Credit: Arxiv
Differentially Private Sparse Linear Regression with Heavy-tailed Responses
- This research paper introduces a method called DP-IHT-H for differentially private sparse linear regression with heavy-tailed responses in high-dimensional settings.
- DP-IHT-H leverages the Huber loss and private iterative hard thresholding to achieve an estimation error bound under the differential privacy model.
- Another method proposed in the paper, DP-IHT-L, further improves the error bound under additional assumptions on the response and achieves better results.
- Experiments conducted on synthetic and real-world datasets show that these methods outperform standard differentially private algorithms designed for regular data.
Read Full Article
4 Likes
Arxiv
1d
53
Image Credit: Arxiv
SAFE: Finding Sparse and Flat Minima to Improve Pruning
- Sparsifying neural networks often leads to performance degradation, and restoring original performance is challenging.
- A new approach called SAFE aims to find subnetworks that are both sparse and flat simultaneously.
- Pruning is formulated as a sparsity-constrained optimization problem with flatness as an objective.
- The SAFE method yields sparse networks with improved generalization performance and shows resilience to noisy data.
Read Full Article
3 Likes
Arxiv
1d
322
Image Credit: Arxiv
Log-Sum-Exponential Estimator for Off-Policy Evaluation and Learning
- A new log-sum-exponential estimator is introduced for off-policy learning and evaluation in logged bandit feedback datasets.
- The estimator addresses challenges such as high variance, low-quality propensity scores, and heavy-tailed reward distributions.
- It demonstrates variance reduction and robustness under heavy-tailed conditions, outperforming traditional inverse propensity score estimators.
- Theoretical analysis and empirical evaluations confirm the practical advantages of the new estimator in off-policy learning and evaluation scenarios.
Read Full Article
19 Likes
Arxiv
1d
202
Image Credit: Arxiv
FREE: Fast and Robust Vision Language Models with Early Exits
- Vision-Language Models (VLMs) have shown remarkable performance improvements in recent years, but their large size can be a challenge for real-world applications with latency concerns.
- To address this issue, a new approach called FREE (Fast and Robust Vision Language Models with Early Exits) proposes employing Early Exit (EE) strategies in VLMs, utilizing adversarial training within a GAN-based framework.
- FREE focuses on input-adaptive inference to increase inference speed with minimal performance drop, training exit classifiers within VLMs to improve accuracy and model robustness while reducing overthinking and mid-crisis instances.
- Experimental results show that FREE speeds up the inference process by more than 1.51x while maintaining comparable performance, with the source code available on GitHub.
Read Full Article
12 Likes
Arxiv
1d
97
Image Credit: Arxiv
Can In-Context Reinforcement Learning Recover From Reward Poisoning Attacks?
- Researchers studied the corruption-robustness of in-context reinforcement learning, specifically focusing on the Decision-Pretrained Transformer (DPT).
- They introduced the Adversarially Trained Decision-Pretrained Transformer (AT-DPT) framework to combat reward poisoning attacks targeting the DPT.
- The AT-DPT framework involves training an attacker to minimize the true reward of the DPT by poisoning environment rewards, while training the DPT model to infer optimal actions from the poisoned data.
- Evaluation results demonstrated that the proposed AT-DPT method outperformed standard bandit algorithms and robust baselines in bandit settings, even against adaptive attackers, and showed robustness in more complex environments beyond bandit scenarios.
Read Full Article
5 Likes
For uninterrupted reading, download the app