A naukri.com initiative
ML News
Arxiv
2d
156
Image Credit: Arxiv
Learning Collective Variables from Time-lagged Generation
- Enhanced sampling techniques aim to observe rare events like state transitions in molecular dynamics simulations by using collective variables (CVs).
- A new framework called TLC has been proposed to learn CVs directly from time-lagged conditions of a generative model, capturing slow dynamic behavior instead of just static information.
- TLC was validated on the Alanine Dipeptide system for two CV-based enhanced sampling tasks, demonstrating superior performance compared to existing machine learning CV discovery methods.
- The study introduces a novel approach to learning CVs that could improve accuracy in observing rare events in molecular dynamics simulations.
Read Full Article
9 Likes
Arxiv
2d
105
Image Credit: Arxiv
HGMP:Heterogeneous Graph Multi-Task Prompt Learning
- A new research paper introduces HGMP, a multi-task prompt learning framework for heterogeneous graph neural networks.
- HGMP aims to improve performance in downstream tasks by reformulating them into a unified graph-level task format, addressing model-task mismatch.
- The framework includes a graph-level contrastive pre-training strategy to leverage heterogeneous information and heterogeneous feature prompts for enhanced performance.
- Experimental results demonstrate that HGMP outperforms baseline methods on various tasks, showcasing its adaptability and effectiveness in the heterogeneous graph domain.
Read Full Article
6 Likes
Arxiv
2d
292
Image Credit: Arxiv
Resolving Token-Space Gradient Conflicts: Token Space Manipulation for Transformer-Based Multi-Task Learning
- Multi-Task Learning (MTL) in shared networks can lead to negative transfer due to differences in task objectives.
- Pre-trained transformers have limitations in adaptability, motivating the development of Dynamic Token Modulation and Expansion (DTME-MTL).
- DTME-MTL addresses gradient conflicts in token space to enhance adaptability and reduce overfitting without duplicating network parameters.
- Experiments show that DTME-MTL offers a scalable and efficient solution for improving transformer-based MTL models.
Read Full Article
17 Likes
Arxiv
2d
58
Image Credit: Arxiv
Uncertainty Quantification for Motor Imagery BCI -- Machine Learning vs. Deep Learning
- Brain-computer interfaces (BCIs) can benefit from uncertainty quantification to enhance accuracy in Motor Imagery classification tasks.
- Research compares uncertainty quantification abilities of established BCI classifiers like CSP-LDA and MDRM against Deep Learning methods.
- CSP-LDA and MDRM-T provide the best uncertainty estimates, while Deep Ensembles and CNNs excel in classifications for Motor Imagery BCI tasks.
- Models showcase the ability to differentiate between easy and difficult classifications, enabling improved accuracy by rejecting ambiguous samples.
Read Full Article
3 Likes
Arxiv
2d
304
Image Credit: Arxiv
Neural Concept Verifier: Scaling Prover-Verifier Games via Concept Encodings
- Prover-Verifier Games (PVGs) are effective for verifiability in nonlinear classification models, but have not been applied to high-dimensional images.
- Concept Bottleneck Models (CBMs) can interpret complex data but rely on low-capacity linear predictors.
- Neural Concept Verifier (NCV) combines PVGs with concept encodings for interpretable, nonlinear classification in high-dimensional settings.
- NCV utilizes concept encodings extracted from raw inputs, outperforming CBMs and pixel-based PVG classifier baselines, showing promise for verifiable AI.
Read Full Article
18 Likes
Arxiv
2d
183
Image Credit: Arxiv
Real-Time Decorrelation-Based Anomaly Detection for Multivariate Time Series
- Anomaly detection (AD) is crucial for identifying deviations in data across various domains.
- Real-time AD is in demand due to the increase in multivariate sensor data from the (Industrial) Internet of Things.
- DAD is a new real-time decorrelation-based anomaly detection method for multivariate time series.
- Experiments show that DAD outperforms state-of-the-art methods in detecting anomalies in high-dimensional data streams.
Read Full Article
11 Likes
Arxiv
2d
121
Image Credit: Arxiv
COALA: Numerically Stable and Efficient Framework for Context-Aware Low-Rank Approximation
- COALA is a new framework for context-aware low-rank approximation in neural networks, aiming to overcome numerical instabilities seen in existing methods.
- Existing methods rely on classical formulas that can lead to degraded approximation quality or numerically singular matrices.
- COALA proposes an inversion-free regularized framework based on stable decompositions to address these limitations.
- The method is capable of handling challenging scenarios like large calibration matrices, nearly singular activation matrices, and insufficient data for unique approximation, providing explicit error bounds.
Read Full Article
7 Likes
Arxiv
2d
124
Image Credit: Arxiv
CHOMET: Conditional Handovers via Meta-Learning
- Handovers (HOs) are crucial in cellular networks for ensuring connectivity to mobile users, but traditional HOs face challenges in complex networks with diverse users and smaller cells.
- To address these challenges, 3GPP introduced conditional handovers (CHOs) which prepare multiple cells for a single user, aiming to increase HO success rates and reduce delays.
- However, CHOs bring new challenges like efficient resource allocation and managing signaling overhead, which require optimization.
- A new framework utilizing meta-learning for CHO optimization within the O-RAN paradigm has shown significant performance improvements, outperforming 3GPP benchmarks by at least 180% in dynamic signal conditions.
Read Full Article
7 Likes
Arxiv
2d
136
Image Credit: Arxiv
Improving Clustering on Occupational Text Data through Dimensionality Reduction
- A study aimed to propose an optimal clustering mechanism for occupations in the O*NET database.
- The study used BERT-based techniques and various clustering approaches to create a map between different definitions of occupations.
- The impact of dimensionality reduction on clustering algorithms' performance metrics was assessed in the study.
- Results improved by utilizing a specialized silhouette approach, potentially aiding individuals in transitioning careers.
Read Full Article
7 Likes
Arxiv
2d
156
Image Credit: Arxiv
Stress Monitoring in Healthcare: An Ensemble Machine Learning Framework Using Wearable Sensor Data
- Healthcare professionals, especially nurses, experience heightened stress levels, accentuated by the COVID-19 crisis.
- A new study presents a novel ensemble machine learning framework for stress monitoring using wearable sensor data.
- The framework addresses data challenges by utilizing a multimodal dataset and advanced ML models like Random Forest and XGBoost.
- The research aims to enhance the development of real-time stress-monitoring systems for healthcare workers' well-being, with future directions including demographic diversity and edge-computing implementations.
Read Full Article
9 Likes
Arxiv
2d
187
Image Credit: Arxiv
Synthetic MC via Biological Transmitters: Therapeutic Modulation of the Gut-Brain Axis
- Synthetic molecular communication (SMC) enables continuous monitoring of biochemical signals in future healthcare systems.
- Closing the loop between sensing and actuation in SMC requires detection and generation of in-body molecular communication signals.
- This paper focuses on using the gut-brain axis (GBA) for therapeutic modulation to indirectly generate signals inside the human body.
- The proposed approach leverages personal health data and machine learning to design more effective GBA modulation-based treatments.
Read Full Article
11 Likes
Arxiv
2d
27
Image Credit: Arxiv
Sparse Causal Discovery with Generative Intervention for Unsupervised Graph Domain Adaptation
- Unsupervised Graph Domain Adaptation (UGDA) aims to achieve effective performance in unlabeled target domains by leveraging labeled source domain graphs despite distribution shifts.
- SLOGAN (Sparse Causal Discovery with Generative Intervention) is a novel approach proposed to address the challenges faced by existing UGDA methods in yielding suboptimal results due to causal-spurious features and global alignment strategies.
- SLOGAN utilizes sparse causal modeling and dynamic intervention mechanisms to achieve stable graph representation transfer by disentangling causal features, compressing domain-dependent correlations, and breaking local spurious couplings through generative intervention.
- Extensive experiments on real-world datasets show that SLOGAN outperforms existing baselines, demonstrating its effectiveness in unsupervised graph domain adaptation.
Read Full Article
1 Like
Arxiv
2d
31
Image Credit: Arxiv
TransformEEG: Towards Improving Model Generalizability in Deep Learning-based EEG Parkinson's Disease Detection
- Electroencephalography (EEG) is being used for early detection of Parkinson's Disease.
- Deep Learning models for EEG-based PD detection lack generalizability due to high inter-subject variability.
- A new model called TransformEEG is introduced, combining Convolutional-Transformer architecture.
- TransformEEG shows improved performance in PD detection with reduced variability and increased consistency.
Read Full Article
1 Like
Arxiv
2d
257
Image Credit: Arxiv
Some Theoretical Results on Layerwise Effective Dimension Oscillations in Finite Width ReLU Networks
- The study analyzes the layerwise effective dimension in fully-connected ReLU networks of finite width.
- Closed-form expressions for the expected rank of the hidden activation matrices for a fixed batch of inputs are derived.
- Main result indicates that the rank deficit decays geometrically with a ratio of 0.3634.
- The oscillatory rank behavior observed is a finite-width phenomenon in random ReLU networks, shedding light on deep network expressivity.
Read Full Article
15 Likes
Arxiv
2d
148
Image Credit: Arxiv
Balancing the Past and Present: A Coordinated Replay Framework for Federated Class-Incremental Learning
- Federated Class Incremental Learning (FCIL) involves processing increasing tasks across multiple clients collaboratively.
- A new method called FedCBDR has been proposed to address class imbalance issues in data replay for FCIL.
- FedCBDR utilizes global coordination for memory construction and adjusts the learning objective to handle imbalances.
- Experimental results show that FedCBDR improves class-wise sampling and generalization, outperforming existing methods by 2%-15% in Top-1 accuracy.
Read Full Article
8 Likes
For uninterrupted reading, download the app