A naukri.com initiative
ML News
Arxiv
1d
151
Image Credit: Arxiv
Dark patterns in e-commerce: a dataset and its baseline evaluations
- Dark patterns are user interface designs in online services that induce users to take unintended actions.
- A dataset for dark pattern detection has been constructed, consisting of 1,818 dark pattern texts from shopping sites.
- State-of-the-art machine learning methods like BERT, RoBERTa, ALBERT, and XLNet have been applied to demonstrate automatic detection accuracy as baselines.
- RoBERTa achieved the highest accuracy of 0.975 in the 5-fold cross-validation.
Read Full Article
9 Likes
Arxiv
1d
116
Image Credit: Arxiv
Interpretable Few-shot Learning with Online Attribute Selection
- Researchers propose an interpretable model for few-shot learning based on human-friendly attributes.
- The model utilizes an online attribute selection mechanism to filter out irrelevant attributes in each episode.
- An automated mechanism is introduced to detect episodes with insufficient available attributes and augment them with learned unknown attributes.
- The proposed method achieves results comparable to black-box few-shot learning models and outperforms other methods in terms of decision alignment with human understanding.
Read Full Article
7 Likes
Arxiv
1d
50
Image Credit: Arxiv
Adversarially Robust Learning with Optimal Transport Regularized Divergences
- Researchers propose a new class of optimal-transport-regularized divergences, D^c, to enhance the adversarial robustness of deep learning models.
- The proposed ARMOR_D methods minimize the maximum expected loss over a D^c-neighborhood of the training data's empirical distribution.
- ARMOR_D allows transportation and re-weighting of adversarial samples, providing enhanced adversarial re-weighting on top of adversarial sample transport.
- The method demonstrates improved performance on CIFAR-10 and CIFAR-100 image recognition, outperforming existing methods against adversarial attacks.
Read Full Article
3 Likes
Arxiv
1d
263
Image Credit: Arxiv
TouchUp-G: Improving Feature Representation through Graph-Centric Finetuning
- The research paper proposes TOUCHUP-G, a method to enhance the node features acquired from Pretrained Models (PMs) for downstream graph tasks.
- TOUCHUP-G is applicable to any downstream graph task, including link prediction in recommender systems.
- It can improve features of any modality like images, texts, and audio.
- TOUCHUP-G effectively shrinks the discrepancy between the graph structure and node features, achieving state-of-the-art results on real-world datasets.
Read Full Article
15 Likes
Arxiv
1d
34
Image Credit: Arxiv
ADMM Algorithms for Residual Network Training: Convergence Analysis and Parallel Implementation
- Researchers propose both serial and parallel proximal (linearized) alternating direction method of multipliers (ADMM) algorithms for training residual neural networks.
- The proposed algorithms mitigate the exploding gradient issue and are suitable for parallel and distributed training through regional updates.
- The algorithms converge at an R-linear (sublinear) rate for both the iteration points and the objective function values.
- Experimental results validate the proposed ADMM algorithms, showing rapid and stable convergence, improved performance, and high computational efficiency.
Read Full Article
2 Likes
Arxiv
1d
108
Image Credit: Arxiv
Integrating Fairness and Model Pruning Through Bi-level Optimization
- Applying fairness criteria to model pruning to address algorithmic biases and social justice concerns.
- Introducing a framework for fair model pruning that optimizes the pruning mask and weight update processes with fairness constraints.
- Demonstrating the superiority of the proposed method in maintaining model fairness, performance, and efficiency compared to mainstream pruning strategies.
- Validating the approach through experiments on various datasets and scenarios.
Read Full Article
6 Likes
Arxiv
1d
236
Image Credit: Arxiv
Model Selection for Inverse Reinforcement Learning via Structural Risk Minimization
- Inverse reinforcement learning (IRL) focuses on selecting the reward function model using structural risk minimization (SRM).
- IRL tackles the trade-off between a simplistic model and one with high complexity to obtain the ideal reward function.
- The SRM framework selects the optimal reward function class that minimizes both estimation error and model complexity.
- Simulations show the algorithm's performance and efficiency in the linear weighted sum setting.
Read Full Article
14 Likes
Arxiv
1d
23
Image Credit: Arxiv
Sample, estimate, aggregate: A recipe for causal discovery foundation models
- Causal discovery, the task of inferring causal structure from data, has the potential to uncover mechanistic insights from biological experiments.
- To address challenges in causal discovery with larger sets of variables and limited data, a foundation model-inspired approach is proposed.
- The approach involves training a supervised model on large-scale, synthetic data to predict causal graphs from summary statistics.
- Experiments show that the model generalizes well, runs on graphs with hundreds of variables in seconds, and is adaptable to different data assumptions.
Read Full Article
1 Like
Arxiv
1d
302
Image Credit: Arxiv
Implicit Bias and Fast Convergence Rates for Self-attention
- Researchers study the implicit bias of self-attention in transformers.
- Convergence of the key-query matrix is possbile with certain conditions.
- Two adaptive step-size strategies, normalized GD and Polyak step-size, are analyzed.
- The findings accelerate parameter convergence and deepen understanding of implicit bias in self-attention.
Read Full Article
18 Likes
Arxiv
1d
314
Image Credit: Arxiv
TimeCMA: Towards LLM-Empowered Multivariate Time Series Forecasting via Cross-Modality Alignment
- Researchers propose TimeCMA, an intuitive framework for Multivariate Time Series Forecasting (MTSF) via cross-modality alignment.
- TimeCMA combines large language models (LLMs) with time series data to achieve improved forecasting performance.
- The framework uses a dual-modality encoding approach to obtain disentangled time series embeddings and robust prompt embeddings.
- TimeCMA outperforms existing methods in MTSF according to extensive experiments on eight real datasets.
Read Full Article
18 Likes
Arxiv
1d
104
Image Credit: Arxiv
Entropy-Reinforced Planning with Large Language Models for Drug Discovery
- A new approach called Entropy-Reinforced Planning (ERP) is proposed for drug discovery using large language models (LLMs).
- LLMs are effective for generating molecules but can result in invalid or suboptimal compounds due to their prior experience.
- ERP enhances the Transformer decoding process by employing an entropy-reinforced planning algorithm, achieving a balance between exploitation and exploration.
- Experimental results show that ERP outperforms the current state-of-the-art algorithm and baselines in multiple benchmarks, including SARS-CoV-2 and human cancer target proteins.
Read Full Article
6 Likes
Arxiv
1d
139
Image Credit: Arxiv
Mosaic of Modalities: A Comprehensive Benchmark for Multimodal Graph Learning
- Graph machine learning has made significant strides in recent years, yet the integration of visual information with graph structure and its potential for improving performance in downstream tasks remains an underexplored area.
- To address this critical gap, the Multimodal Graph Benchmark (MM-GRAPH) is introduced as a comprehensive evaluation framework for multimodal graph learning, incorporating both visual and textual information into graph learning tasks.
- MM-GRAPH consists of seven diverse datasets, designed to assess algorithms across different tasks in real-world scenarios, featuring rich multimodal node attributes including visual data.
- Through an extensive empirical study, valuable insights into the challenges and opportunities of integrating visual data into graph learning are provided.
Read Full Article
8 Likes
Arxiv
1d
368
Image Credit: Arxiv
Sustainable techniques to improve Data Quality for training image-based explanatory models for Recommender Systems
- A new research study explores sustainable techniques to improve the data quality for training image-based explanatory models for Recommender Systems.
- Current approaches in this domain often suffer from limitations due to sparse and noisy training data.
- To address this, the study introduces three novel strategies for training data quality enhancement, including reliable negative training example selection, transform-based data augmentation, and text-to-image generative-based data augmentation.
- Integration of these strategies in explainability models resulted in a 5% performance increase without compromising long-term sustainability.
Read Full Article
22 Likes
Arxiv
1d
50
Image Credit: Arxiv
Backdoor Graph Condensation
- Graph condensation is a technique to improve the training efficiency for graph neural networks (GNNs).
- Researchers have introduced an effective backdoor attack called BGC against graph condensation.
- BGC targets representative nodes for poisoning and achieves a high attack success rate.
- The attack proves resilient against multiple defense methods.
Read Full Article
3 Likes
Arxiv
1d
15
Image Credit: Arxiv
Krait: A Backdoor Attack Against Graph Prompt Tuning
- Graph prompt tuning has emerged as a promising paradigm for transferring general graph knowledge from pre-trained models to downstream tasks.
- A backdoor attack known as Krait has been introduced, which disguises benign graph prompts to evade detection.
- Krait efficiently embeds triggers to a small fraction of training nodes, achieving high attack success rates without sacrificing clean accuracy.
- The study analyzes how Krait can evade both classical and state-of-the-art defenses and provides insights for detecting and mitigating such attacks.
Read Full Article
Like
For uninterrupted reading, download the app