A naukri.com initiative
ML News
Arxiv
6d
12
Image Credit: Arxiv
Entropic bounds for conditionally Gaussian vectors and applications to neural networks
- Using entropic inequalities from information theory, new bounds on the total variation and 2-Wasserstein distances between conditionally Gaussian and Gaussian laws are provided.
- The results are applied to quantify the convergence speed of a randomly initialized fully connected neural network and its derivatives to Gaussian distributions.
- The findings improve and extend previous research studies on the subject.
- Cumulant estimates and activation function assumptions play a crucial role in the results.
Read Full Article
Like
Arxiv
6d
260
Image Credit: Arxiv
Single View Garment Reconstruction Using Diffusion Mapping Via Pattern Coordinates
- Reconstructing 3D clothed humans from images is important for virtual try-on, avatar creation, and mixed reality.
- Accurate reconstruction of loose-fitting garment geometry is a challenge.
- A novel method for high-fidelity 3D garment reconstruction from single images is proposed.
- The approach combines Implicit Sewing Patterns (ISP) with a generative diffusion model for learning garment shape priors and establishes correspondences between 2D image pixels, UV pattern coordinates, and 3D geometry.
Read Full Article
15 Likes
Arxiv
6d
409
Image Credit: Arxiv
FocalLens: Instruction Tuning Enables Zero-Shot Conditional Image Representations
- Visual understanding is inherently contextual, and what we focus on in an image depends on the task at hand.
- FocalLens is a conditional visual encoding method that produces different representations for the same image based on the context of interest, expressed through natural language instructions.
- FocalLens outperforms generic visual encoders by better highlighting the visual features of interest.
- FocalLens improves performance on various downstream tasks, including image-image retrieval, image classification, and image-text retrieval.
Read Full Article
24 Likes
Arxiv
6d
222
Image Credit: Arxiv
An Empirical Investigation of Reconstruction-Based Models for Seizure Prediction from ECG Signals
- Epileptic seizures can be predicted using ECG signals as an alternative to EEG signals.
- A novel method based on detecting anomalies in ECG signals during their reconstruction was proposed.
- The method achieved specificity of 99.16%, accuracy of 76.05%, and a low false positive rate.
- ECG-based seizure prediction has the potential to be a patient-friendly alternative to EEG-based methods.
Read Full Article
13 Likes
Arxiv
6d
378
Image Credit: Arxiv
MixDiT: Accelerating Image Diffusion Transformer Inference with Mixed-Precision MX Quantization
- MixDiT is an algorithm-hardware co-designed acceleration solution proposed to address the compute-intensive nature and long latency of Diffusion Transformer (DiT) inferencing.
- MixDiT utilizes mixed Microscaling (MX) formats to quantize DiT activation values, selectively applying higher precision to magnitude-based outliers to achieve low-precision quantization with high accuracy and substantial speedup.
- A MixDiT accelerator is designed to enable precision-flexible multiplications and efficient MX precision conversions, resulting in a speedup of 2.10-5.32 times over RTX 3090 in experimental results.
- MixDiT achieves this speedup without any loss in Fréchet Inception Distance (FID).
Read Full Article
22 Likes
Arxiv
6d
19
Image Credit: Arxiv
Belief States for Cooperative Multi-Agent Reinforcement Learning under Partial Observability
- Reinforcement learning in partially observable environments is challenging, especially in multi-agent settings.
- The authors propose using learned beliefs on the underlying system state to overcome these challenges.
- Belief states are pre-trained in a self-supervised fashion and used in a state-based reinforcement learning algorithm.
- The proposed method simplifies learning tasks, improves convergence speed, and final performance.
Read Full Article
Like
Arxiv
6d
148
Image Credit: Arxiv
Standardization of Weighted Ranking Correlation Coefficients
- A relevant problem in statistics is defining the correlation of two rankings of a list of items.
- Weighted versions of the original Spearman and Kendall coefficients have emerged to address the greater importance of top ranks.
- This standardization function maps a correlation ranking coefficient to a standard form with zero expected value.
- The proposed standardization maintains the relevant statistical properties of the original coefficient.
Read Full Article
8 Likes
Arxiv
6d
250
Image Credit: Arxiv
Artifact detection and localization in single-channel mobile EEG for sleep research using deep learning and attention mechanisms
- Artifacts in the electroencephalogram (EEG) degrade signal quality and impact the analysis of brain activity.
- A convolutional neural network (CNN) model incorporating a convolutional block attention module (CNN-CBAM) is proposed to detect and identify the location of artifacts in sleep EEG with attention maps.
- The CNN-CBAM model achieved high performance with the highest area under the receiver operating characteristic curve (0.88), sensitivity (0.81), and specificity (0.86) when compared to other approaches.
- This work demonstrates the feasibility of automating the detection and localization of artifacts in wearable sleep EEG.
Read Full Article
15 Likes
Arxiv
6d
320
Image Credit: Arxiv
AstroLLaVA: towards the unification of astronomical data and natural language
- AstroLLaVA is a vision language model for astronomy that enables interaction with astronomical imagery through natural dialogue.
- It is fine-tuned on a diverse dataset of approximately 30k images with captions and question-answer pairs.
- AstroLLaVA is capable of answering open-ended questions about astronomical concepts depicted visually.
- The model weights, code, and training set of AstroLLaVA are released for further open-source work in the field of astronomy.
Read Full Article
19 Likes
Arxiv
6d
324
Image Credit: Arxiv
Boosting multi-demographic federated learning for chest x-ray analysis using general-purpose self-supervised representations
- Reliable artificial intelligence (AI) models for medical image analysis often depend on large and diverse labeled datasets.
- Federated learning (FL) struggles in highly non-independent and identically distributed (non-IID) settings, where institutions with more representative data may experience degraded performance.
- Large-scale FL studies have been limited to adult datasets, neglecting the unique challenges posed by pediatric data, which introduces additional non-IID variability.
- Using transfer learning from general-purpose self-supervised image representations, the study analyzed adult and pediatric chest radiographs and found that FL improved performance for smaller adult datasets but degraded performance for larger datasets and pediatric cases.
Read Full Article
19 Likes
Arxiv
6d
329
Image Credit: Arxiv
Ready, Bid, Go! On-Demand Delivery Using Fleets of Drones with Unknown, Heterogeneous Energy Storage Constraints
- Unmanned Aerial Vehicles (UAVs) are expected to transform logistics, reducing delivery time, costs, and emissions.
- This study addresses on-demand delivery using fleets of UAVs.
- The UAVs have heterogeneous, unknown energy storage capacities and no knowledge of energy consumption models.
- The proposed strategy combines auction-based task allocation with online learning, resulting in improved delivery times and successful order fulfillment.
Read Full Article
19 Likes
Arxiv
6d
370
Image Credit: Arxiv
On Background Bias of Post-Hoc Concept Embeddings in Computer Vision DNNs
- The study investigates biases in post-hoc approaches of concept-based explainable AI (C-XAI) in deep neural networks (DNNs).
- Existing approaches capture background biases, leading to performance degradation in certain scenarios.
- The study validates this on >50 concepts from 2 datasets and 7 DNN architectures.
- Even low-cost setups can provide valuable insights and improved background robustness.
Read Full Article
22 Likes
Arxiv
6d
379
Image Credit: Arxiv
Neural Fidelity Calibration for Informative Sim-to-Real Adaptation
- Neural Fidelity Calibration (NFC) is a novel framework that calibrates simulator physical coefficients and residual fidelity domains online during robot execution.
- It fine-tunes the pretrained policy only under anomalous scenarios and builds sequential NFC online with the pretrained NFC's proposal prior.
- NFC leverages optimistic exploration to enable hallucinated policy optimization when uncertainty is high and may degrade policy improvement.
- The framework achieves superior simulator calibration precision and demonstrates robust robot navigation under challenging real-world conditions.
Read Full Article
22 Likes
Arxiv
6d
238
Image Credit: Arxiv
Gradient Descent Robustly Learns the Intrinsic Dimension of Data in Training Convolutional Neural Networks
- Modern neural networks are usually highly over-parameterized.
- This work studies the rank of convolutional neural networks (CNNs) trained by gradient descent.
- CNNs trained with gradient descent are found to be robust to image background noises.
- Theoretical case study and experiments on synthetic and real datasets support the claim.
Read Full Article
14 Likes
Arxiv
6d
29
Image Credit: Arxiv
Genius: A Generalizable and Purely Unsupervised Self-Training Framework For Advanced Reasoning
- Advancing LLM reasoning skills has captivated wide interest.
- Current post-training techniques rely on supervisory signals, which face scalability and high annotation costs.
- Genius is a generalizable and purely unsupervised self-training framework for enhancing LLM reasoning.
- It introduces stepwise foresight re-sampling and advantage-calibrated optimization to optimize LLM without external supervision.
Read Full Article
1 Like
For uninterrupted reading, download the app