A naukri.com initiative
ML News
Arxiv
3d
384
Image Credit: Arxiv
ATHEENA: A Toolflow for Hardware Early-Exit Network Automation
- The paper introduces ATHEENA, a Toolflow for Hardware Early-Exit Network Automation.
- ATHEENA leverages the probability of samples exiting early to optimize resource allocation in FPGA networks.
- The toolflow uses data-flow model and Design Space Exploration to improve throughput and reduce area while maintaining accuracy.
- Experimental results show throughput increase and resource reduction in comparison to baseline network implementations.
Read Full Article
23 Likes
Arxiv
3d
40
Image Credit: Arxiv
BotArtist: Generic approach for bot detection in Twitter via semi-automatic machine learning pipeline
- A research paper introduces a semi-automatic machine learning pipeline (SAMLP) called BotArtist for bot detection on Twitter.
- SAMLP leverages nine publicly available datasets to train the BotArtist model.
- BotArtist outperforms existing Twitter bot detection methods by almost 10% in terms of F1-score, achieving an average score of 83.19% and 68.5% over specific and general approaches, respectively.
- The research provides one of the largest labeled Twitter bot datasets, containing features and BotArtist predictions for 10,929,533 Twitter user profiles.
Read Full Article
2 Likes
Arxiv
3d
124
Image Credit: Arxiv
Tailoring Adversarial Attacks on Deep Neural Networks for Targeted Class Manipulation Using DeepFool Algorithm
- Researchers have developed the Enhanced Targeted DeepFool (ET DeepFool) algorithm for tailoring adversarial attacks on deep neural networks.
- The algorithm allows for the specification of desired misclassification targets and incorporates a configurable minimum confidence score.
- Preliminary outcomes suggest that certain models, including AlexNet and the Vision Transformer, exhibit robustness to the manipulations enabled by ET DeepFool.
- The code for the algorithm is available on GitHub at https://github.com/FazleLabib/et_deepfool.
Read Full Article
7 Likes
Arxiv
3d
180
Image Credit: Arxiv
Toward 6G Native-AI Network: Foundation Model based Cloud-Edge-End Collaboration Framework
- Future wireless communication networks are moving towards intelligent, immersive experiences based on multi-agent collaboration and 6G native artificial intelligence.
- This paper analyzes the challenges of achieving 6G native AI and proposes a framework based on foundation models to address those challenges.
- The framework includes an integration method for expert knowledge, customization for foundation models, and a novel operational paradigm.
- A practical use case of the framework is applied to achieve the maximum sum rate in a cell-free massive MIMO system.
Read Full Article
10 Likes
Arxiv
3d
312
Image Credit: Arxiv
PCDP-SGD: Improving the Convergence of Differentially Private SGD via Projection in Advance
- The paper proposes a framework called PCDP-SGD to improve the convergence of Differentially Private Stochastic Gradient Descent (DP-SGD).
- PCDP-SGD compresses redundant gradient norms and preserves crucial top gradient components through a projection operation before gradient clipping.
- The framework is extended to differential privacy federated learning (DPFL) to tackle data heterogeneity and achieve efficient communication.
- Experimental results show that PCDP-SGD achieves higher accuracy compared to other DP-SGD variants and outperforms current federated learning frameworks in computer vision tasks.
Read Full Article
18 Likes
Arxiv
3d
48
Image Credit: Arxiv
Artificial Neural Network for Estimation of Physical Parameters of Sea Water using LiDAR Waveforms
- LiDAR is a remote sensing technology that uses laser beams to measure distances and create 3D representations of objects and environments.
- A research proposes a novel solution using neural networks to estimate physical parameters, such as depth, attenuation coefficient, and bottom reflectance, in LiDAR data analysis.
- Current techniques for parameter estimation suffer from limitations in accuracy, and there is no established method for predicting bottom reflectance.
- The proposed solution successfully learned the inversion model and achieved promising results when tested on real LiDAR data.
Read Full Article
2 Likes
Arxiv
3d
136
Image Credit: Arxiv
Boosting Column Generation with Graph Neural Networks for Joint Rider Trip Planning and Crew Shift Scheduling
- This study addresses the complexities of service scheduling by jointly optimizing rider trip planning and crew scheduling for a dynamic mobility service.
- The paper introduces the Joint Rider Trip Planning and Crew Shift Scheduling Problem (JRTPCSSP) and a solution method called Attention and Gated GNN-Informed Column Generation (AGGNNI-CG).
- AGGNNI-CG hybridizes column generation and machine learning to obtain near-optimal solutions with real-life constraints.
- With its graph neural network and attention mechanism, AGGNNI-CG significantly improves service quality and produces substantial improvements compared to baseline approaches.
Read Full Article
8 Likes
Arxiv
3d
32
Image Credit: Arxiv
COMPASS: Computational Mapping of Patient-Therapist Alliance Strategies with Language Modeling
- The therapeutic working alliance is a critical predictor of psychotherapy success.
- COMPASS is a framework that directly infers the therapeutic working alliance from natural language used in psychotherapy sessions.
- Using large language models, COMPASS analyzes session transcripts and maps them to distributed representations.
- COMPASS enhances understanding of therapeutic interactions and provides actionable insights to improve the effectiveness of psychotherapy.
Read Full Article
1 Like
Arxiv
3d
84
Image Credit: Arxiv
Q-Newton: Hybrid Quantum-Classical Scheduling for Accelerating Neural Network Training with Newton's Gradient Descent
- Optimization techniques in deep learning are predominantly led by first-order gradient methodologies like SGD, but second-order optimization methods like Newton's GD can greatly benefit neural network training.
- Matrix inversion is a major bottleneck for Newton's GD, with a time complexity of O(N^3).
- The use of quantum linear solver algorithms (QLSAs) presents a promising approach to accelerate matrix inversion with exponentially reduced time complexity of O(d * κ * log(N * κ / ε)).
- Q-Newton is a hybrid quantum-classical scheduler proposed to accelerate neural network training with Newton's GD by coordinating between quantum and classical linear solvers.
Read Full Article
5 Likes
Arxiv
3d
360
Image Credit: Arxiv
Contribute to balance, wire in accordance: Emergence of backpropagation from a simple, bio-plausible neuroplasticity rule
- A novel neuroplasticity rule has been proposed for implementing backpropagation (BP) in the brain.
- The rule is based on maintaining the balance of excitatory and inhibitory inputs, as well as on retrograde signaling.
- It operates over three progressively slower timescales: neural firing, retrograde signaling, and neural plasticity.
- Simulations on artificial neural networks show that this rule induces varying community structures in networks.
Read Full Article
21 Likes
Arxiv
3d
196
Image Credit: Arxiv
SnatchML: Hijacking ML models without Training Access
- Model hijacking can cause significant accountability and security risks.
- SnatchML is a training-free model hijacking attack that targets inference-time.
- SnatchML leverages the over-parameterization of ML models to infer different tasks.
- The study proposes countermeasures to mitigate the risks of model hijacking.
Read Full Article
11 Likes
Arxiv
3d
248
Image Credit: Arxiv
Beyond Statistical Estimation: Differentially Private Individual Computation via Shuffling
- A new paradigm called Private Individual Computation (PIC) is introduced in this paper.
- The shuffle model of differential privacy is expanded to support a broader range of permutation-equivariant computations.
- PIC enables personalized outputs while preserving privacy, utilizing privacy amplification through shuffling.
- The proposed PIC protocol and the Minkowski randomizer demonstrate superior utility compared to existing solutions.
Read Full Article
14 Likes
Arxiv
3d
264
Image Credit: Arxiv
Towards A Generalizable Pathology Foundation Model via Unified Knowledge Distillation
- Foundation models pretrained on large-scale datasets are revolutionizing the field of computational pathology (CPath).
- Existing foundation models excel at certain clinical task types but struggle to handle the full breadth of tasks in the field.
- To improve the generalization of pathology foundation models, a unified knowledge distillation framework is proposed, combining expert and self-knowledge distillation.
- The Generalizable Pathology Foundation Model (GPFM) achieved an impressive average rank of 1.6 in six distinct clinical task types.
Read Full Article
15 Likes
Arxiv
3d
264
Image Credit: Arxiv
Cybench: A Framework for Evaluating Cybersecurity Capabilities and Risks of Language Models
- Cybench is a framework introduced for evaluating cybersecurity capabilities and risks of language model agents for autonomous vulnerability identification and exploit execution.
- It includes 40 professional-level Capture the Flag (CTF) tasks from 4 distinct competitions, providing a wide range of difficulties.
- By evaluating various language models, including GPT-4o and Claude 3.5 Sonnet, it was found that models could successfully solve tasks that took human teams hours to complete.
- The framework and all related code and data are publicly available at https://cybench.github.io.
Read Full Article
15 Likes
Arxiv
3d
180
Image Credit: Arxiv
Integrated Brain Connectivity Analysis with fMRI, DTI, and sMRI Powered by Interpretable Graph Neural Networks
- Researchers have developed an approach to integrate and interpret multimodal neuroimaging data within a cohesive analytical framework.
- The study combines functional magnetic resonance imaging (fMRI), diffusion tensor imaging (DTI), and structural MRI (sMRI) to understand brain connectivity and anatomical characteristics.
- Using a masking strategy, the model differentially weighs neural connections to enhance interpretability at the connectivity level.
- The approach was applied to the Human Connectome Project's Development study, revealing associations between multimodal imaging and cognitive functions in youth.
Read Full Article
10 Likes
For uninterrupted reading, download the app