A naukri.com initiative
ML News
Arxiv
2d
359
Image Credit: Arxiv
Electric Bus Charging Schedules Relying on Real Data-Driven Targets Based on Hierarchical Deep Reinforcement Learning
- The study focuses on the charging scheduling problem for Electric Buses (EBs) using Deep Reinforcement Learning (DRL).
- Hierarchical DRL (HDRL) is proposed to tackle the challenge of long-range multi-phase planning with sparse rewards.
- The algorithm involves Hierarchical Double Deep Q-Network (HDDQN)-Hindsight Experience Replay (HER) for decision-making at different temporal resolutions.
- Numerical experiments with real-world data have been conducted to assess the algorithm's performance.
Read Full Article
21 Likes
Arxiv
2d
349
Image Credit: Arxiv
Cutting Through Privacy: A Hyperplane-Based Data Reconstruction Attack in Federated Learning
- A new data reconstruction attack called Hyperplane-Based Data Reconstruction Attack has been introduced in Federated Learning (FL).
- This attack overcomes limitations of existing data reconstruction attacks by leveraging a geometric perspective on fully connected layers.
- The method enables the perfect recovery of arbitrarily large data batches in classification tasks without prior knowledge of clients' data.
- Experiments on image and tabular datasets show that this attack outperforms existing methods and achieves perfect reconstruction of data batches two orders of magnitude larger than the state of the art.
Read Full Article
21 Likes
Arxiv
2d
368
Image Credit: Arxiv
Optimizing Electric Bus Charging Scheduling with Uncertainties Using Hierarchical Deep Reinforcement Learning
- The article discusses the challenges faced in optimizing Electric Bus (EB) charging schedules due to uncertainties in travel time, energy consumption, and fluctuating electricity prices.
- A solution proposed in the paper is the use of Hierarchical Deep Reinforcement Learning (HDRL) to reformulate the Markov Decision Process (MDP) into two augmented MDPs for efficient decision-making across multiple time scales.
- The novel HDRL algorithm introduced, called Double Actor-Critic Multi-Agent Proximal Policy Optimization Enhancement (DAC-MAPPO-E), addresses scalability challenges for large EB fleets by redesigning the decentralized actor network and incorporating the Multi-Agent Proximal Policy Optimization (MAPPO) algorithm.
- Extensive experiments with real-world data support the superior performance and scalability of DAC-MAPPO-E in optimizing EB fleet charging schedules.
Read Full Article
22 Likes
Arxiv
2d
365
Image Credit: Arxiv
Defending the Edge: Representative-Attention for Mitigating Backdoor Attacks in Federated Learning
- Federated learning (FL) improves privacy and reduces communication cost for edge clients by enabling distributed model training at the edge.
- The diverse nature of edge devices leads to non-IID data, posing challenges in detecting backdoor attacks in FL.
- A defense mechanism called FeRA has been introduced, utilizing representative-attention to differentiate between benign and malicious clients.
- FeRA calculates anomaly scores based on reconstruction errors, effectively identifying clients with activations deviating from the consensus, reducing backdoor attack success rates.
Read Full Article
21 Likes
Arxiv
2d
40
Image Credit: Arxiv
Negative Metric Learning for Graphs
- Graph contrastive learning often suffers from false negatives, impacting downstream task performance.
- A novel Negative Metric Learning (NML) enhanced GCL (NML-GCL) is proposed to address the false negative issue.
- NML-GCL employs a learnable Negative Metric Network (NMN) to create a negative metric space for better distinction between false negatives and true negatives.
- A joint training scheme with bi-level optimization objective is suggested to optimize the encoder and the negative metric network effectively, showing superior performance in experiments.
Read Full Article
2 Likes
Arxiv
2d
198
Image Credit: Arxiv
A Representation Learning Approach to Feature Drift Detection in Wireless Networks
- AI is expected to play a vital role in future wireless networks, but changes in feature distribution can impact AI models' performance negatively.
- A new method called ALERT has been introduced to detect feature distribution changes, prompting model re-training in scenarios like wireless fingerprinting and link anomaly detection.
- ALERT comprises representation learning, statistical testing, and utility assessment components, utilizing MLP for representation learning and Kolmogorov-Smirnov tests for statistical testing.
- The proposed ALERT method has shown superior performance compared to ten standard drift detection methods in two wireless network use cases.
Read Full Article
11 Likes
Arxiv
2d
91
Image Credit: Arxiv
Emergence of Structure in Ensembles of Random Neural Networks
- Randomness in systems often leads to emergent global behaviors transitioning from disorder to organization.
- A theoretical model for studying collective behaviors in ensembles of random classifiers has been introduced.
- Ensembles weighted through Gibbs measure at a finite temperature parameter show optimal classification based on loss.
- Experiments on MNIST dataset highlight the universal nature of observed behavior in high-quality, noiseless data.
Read Full Article
5 Likes
Arxiv
2d
353
Image Credit: Arxiv
An Introduction to Discrete Variational Autoencoders
- Variational Autoencoders (VAEs) are widely used for probabilistic unsupervised learning with neural networks.
- Recently, there has been a growing interest in discrete latent spaces, particularly for data modalities like text.
- A tutorial has been provided on discrete variational autoencoders, focusing on VAEs with latent variables following a categorical distribution.
- The tutorial covers the theoretical foundation, practical aspects, training guidelines, and includes an example implementation.
Read Full Article
21 Likes
Arxiv
2d
122
Image Credit: Arxiv
FactsR: A Safer Method for Producing High Quality Healthcare Documentation
- AI-scribing solutions in healthcare often rely on one-shot or few-shot prompts, leading to long notes and potential errors.
- The FactsR method introduced aims to extract clinical information in real-time during consultations for generating more accurate and concise notes.
- FactsR involves clinicians in the note generation process, improving patient safety and enabling new use cases in real-time decision support.
Read Full Article
7 Likes
Arxiv
2d
194
Image Credit: Arxiv
Two-Stage Generative Model for Intracranial Aneurysm Meshes with Morphological Marker Conditioning
- A generative model for the mesh geometry of intracranial aneurysms (IA) called AneuG has been proposed to aid in predicting blood flow forces in real time.
- Existing methods struggle to capture realistic IA features and ignore the relationship between IA pouches and parent vessels, limiting their physiological realism.
- AneuG is a two-stage Variational Autoencoder (VAE)-based IA mesh generator that generates both aneurysm pouch shapes and parent vessels by utilizing Graph Harmonic Deformation (GHD) tokens.
- The IA shape generation by AneuG can be conditioned to have specific morphological measurements, which can be beneficial for understanding shape variations in clinical studies and simulating the effects on fluid dynamics.
Read Full Article
11 Likes
Arxiv
2d
365
Image Credit: Arxiv
Decomposed Inductive Procedure Learning: Learning Academic Tasks with Human-Like Data Efficiency
- Human learning relies on specialization with distinct cognitive mechanisms, while most neural networks rely on gradient descent over an objective function.
- Research investigates if human learners' faster learning with fewer examples compared to data-driven deep learning is due to using multiple specialized mechanisms in combination.
- A study on inductive human learning simulations in tutoring environments shows that decomposing learning into multiple mechanisms significantly improves data efficiency, aligning it with human learning.
- Efforts to improve machine learning efficiency should consider integrating multiple specialized learning mechanisms to bridge the efficiency gap between data-driven approaches and human learning.
Read Full Article
21 Likes
Arxiv
2d
361
Image Credit: Arxiv
The Power of Random Features and the Limits of Distribution-Free Gradient Descent
- A study explores the connection between gradient-based optimization of parametric models like neural networks and optimization of linear combinations of random features.
- The main finding suggests that if a parametric model can be learned using mini-batch stochastic gradient descent without requiring assumptions about data distribution, then the target function can be approximated using a polynomial-sized combination of random features with high probability.
- The size of the combination of random features depends on the number of gradient steps and numerical precision utilized in the bSGD process.
- The study highlights the limitations of distribution-free learning in neural networks trained by gradient descent and emphasizes the importance of making assumptions about data distributions in practical scenarios.
Read Full Article
21 Likes
Arxiv
2d
353
Image Credit: Arxiv
Learning to Think: Information-Theoretic Reinforcement Fine-Tuning for LLMs
- Large language models (LLMs) excel at complex tasks due to their improved reasoning abilities but often overlook the trade-off between reasoning effectiveness and computational efficiency.
- To address this issue, a new framework called Learning to Think (L2T) has been proposed, which is an information-theoretic reinforcement fine-tuning approach for LLMs.
- L2T treats each query-response interaction as a hierarchical session of multiple episodes and uses a universal dense process reward to optimize reasoning with fewer tokens, without the need for additional annotations.
- Theoretical analyses and empirical results demonstrate that L2T optimizes the model through reinforcement learning, leading to improved reasoning effectiveness and efficiency across various tasks and models.
Read Full Article
21 Likes
Arxiv
2d
242
Image Credit: Arxiv
Score-based diffusion nowcasting of GOES imagery
- Machine learning method called score-based diffusion is explored for nowcasting clouds and precipitation in a zero to three hour forecast.
- Three main types of diffusion models were experimented with: standard score-based diffusion model, residual correction diffusion model, and latent diffusion model.
- Results show that these diffusion models can advect existing clouds, generate and decay clouds, and even predict convective initiation.
- The best performing diffusion model was the CorrDiff approach, outperforming traditional U-Net and persistence forecast in root mean squared error.
Read Full Article
14 Likes
Arxiv
2d
218
Image Credit: Arxiv
Identification and Optimal Nonlinear Control of Turbojet Engine Using Koopman Eigenfunction Model
- Gas turbine engines are complex nonlinear dynamical systems, making it challenging to derive physics-based models.
- Conventional experimental methods for deriving component-level and locally linear parameter-varying models have limitations, addressed through data-driven identification techniques.
- Rotor dynamics were estimated using sparse identification of nonlinear dynamics, followed by mapping into an optimally constructed Koopman eigenfunction space.
- A globally optimal nonlinear feedback controller based on the Koopman model outperformed other benchmark controllers in reference tracking and disturbance rejection, showcasing superior performance under varying flight conditions.
Read Full Article
13 Likes
For uninterrupted reading, download the app