A naukri.com initiative
ML News
Arxiv
3d
160
Image Credit: Arxiv
ZeroMimic: Distilling Robotic Manipulation Skills from Web Videos
- ZeroMimic is a system that generates immediately deployable image goal-conditioned skill policies for several common categories of manipulation tasks.
- It uses large pre-recorded human video datasets demonstrating manipulation skills to distill useful robotic skill policies.
- ZeroMimic is trained on the EpicKitchens dataset and can handle diverse objects and unseen task setups.
- Software and policy checkpoints of ZeroMimic skill policies are released for plug-and-play reuse on other task setups and robots.
Read Full Article
9 Likes
Arxiv
3d
4
Image Credit: Arxiv
Feature learning from non-Gaussian inputs: the case of Independent Component Analysis in high dimensions
- Deep neural networks learn structured features from complex, non-Gaussian inputs, but the mechanisms behind this process remain poorly understood.
- The first-layer filters learnt by deep convolutional neural networks from natural images resemble those learnt by independent component analysis (ICA).
- FastICA, a popular ICA algorithm, requires a large number of samples to recover non-Gaussian directions from high dimensional inputs.
- Vanilla online stochastic gradient descent (SGD) outperforms FastICA in feature learning.
Read Full Article
Like
Arxiv
3d
40
Image Credit: Arxiv
Certified Approximate Reachability (CARe): Formal Error Bounds on Deep Learning of Reachable Sets
- Recent approaches to leveraging deep learning for computing reachable sets of continuous-time dynamical systems have gained popularity over traditional level-set methods.
- The introduction of an epsilon-approximate Hamilton-Jacobi Partial Differential Equation (HJ-PDE) establishes a relationship between training loss and accuracy of the true reachable set.
- Satisfiability Modulo Theories (SMT) solvers are used to bound the residual error of the HJ-based loss function, allowing for formal certification of the approximation.
- Certified Approximate Reachability (CARe) is the first approach to provide soundness guarantees on learned reachable sets of continuous dynamical systems.
Read Full Article
2 Likes
Arxiv
3d
184
Image Credit: Arxiv
Detecting Localized Density Anomalies in Multivariate Data via Coin-Flip Statistics
- Detecting localized density differences in multivariate data is a crucial task in computational science.
- Introducing EagleEye, an anomaly detection method for identifying local density anomalies in multivariate datasets.
- Anomalies are detected by modelling the ordered sequence of each point's neighbors' membership labels as a coin-flipping process.
- EagleEye successfully detects anomalies in synthetic and real-world datasets, including particle decay events in Large Hadron Collider data and changes in temperature fields.
Read Full Article
11 Likes
Arxiv
3d
276
Image Credit: Arxiv
Machine Learning-assisted High-speed Combinatorial Optimization with Ising Machines for Dynamically Changing Problems
- Quantum or quantum-inspired Ising machines have shown promise in solving combinatorial optimization problems quickly.
- Real-world applications require solving dynamically changing problems, posing challenges for Ising machines.
- Researchers have developed a method using embedded Ising machines to solve diverse problems at high speed.
- The approach involves customizing the algorithm, circuit architecture, and utilizing a machine learning model for parameter estimation.
Read Full Article
16 Likes
Arxiv
3d
12
Image Credit: Arxiv
The more the merrier: logical and multistage processors in credit scoring
- This paper focuses on the application of fair machine learning (ML) in the context of credit scoring in finance.
- The paper introduces logical processors (LP), a new technique for addressing the application of fairness methods on multiple sensitive variables.
- The paper also explores multistage processors (MP) to determine if combining fairness methods can enhance fairness and accuracy in credit scoring.
- The results indicate that logical processors are suitable for handling multiple sensitive variables, and multistage processors can improve existing methods.
Read Full Article
Like
Arxiv
3d
228
Image Credit: Arxiv
Crossmodal Knowledge Distillation with WordNet-Relaxed Text Embeddings for Robust Image Classification
- Crossmodal knowledge distillation (KD) enhances a unimodal student using a multimodal teacher model.
- A multi-teacher crossmodal KD framework is proposed, integrating CLIP image embeddings with WordNet-relaxed text embeddings.
- This approach reduces label leakage and introduces more diverse textual cues for improved knowledge transfer.
- The method achieves state-of-the-art or second-best results on six public datasets, demonstrating its effectiveness in crossmodal KD.
Read Full Article
13 Likes
Arxiv
3d
236
Image Credit: Arxiv
AutoML Algorithms for Online Generalized Additive Model Selection: Application to Electricity Demand Forecasting
- Electricity demand forecasting is essential to prevent blackouts and ensure supply meets demand.
- A combination of a Generalized Additive Models (GAM) with a State-Space model leads to adaptive (online) forecasting.
- The formula and adaptation parameters of the GAM need to be fixed before model training, affecting predictive performance.
- The DRAGON package is used to optimize these parameters in automated online generalized additive model selection.
Read Full Article
14 Likes
Arxiv
3d
28
Image Credit: Arxiv
Artificial Conversations, Real Results: Fostering Language Detection with Synthetic Data
- Researchers have proposed a pipeline for generating synthetic data and investigating the factors that influence the validity of the data.
- They explored the use of Large Language Models (LLMs) to generate synthetic datasets for language detection.
- The study focused on inclusive language detection in Italian job advertisements.
- Results show that the fine-tuned models trained on synthetic data performed better than other models on both real and synthetic test datasets.
Read Full Article
1 Like
Arxiv
3d
144
Image Credit: Arxiv
HACTS: a Human-As-Copilot Teleoperation System for Robot Learning
- HACTS (Human-As-Copilot Teleoperation System) is a novel system that enables bilateral, real-time joint synchronization between a robot arm and teleoperation hardware.
- The system allows the human copilot to intervene seamlessly while collecting action-correction data for future learning.
- HACTS is implemented using 3D-printed components and low-cost, off-the-shelf motors, making it accessible and scalable.
- The experiments show that HACTS significantly enhances performance in imitation learning (IL) and reinforcement learning (RL) tasks, boosting IL recovery capabilities and data efficiency, and facilitating human-in-the-loop RL.
Read Full Article
8 Likes
Arxiv
3d
148
Image Credit: Arxiv
From Colors to Classes: Emergence of Concepts in Vision Transformers
- Vision Transformers (ViTs) are powerful in computer vision tasks due to their representation capabilities.
- A layer-wise analysis of ViTs using neuron labeling reveals that concepts encoded in ViTs become more complex throughout the network.
- Early layers primarily encode basic features like colors and textures, while later layers represent more specific classes, such as objects and animals.
- Different pretraining strategies influence the quantity and category of encoded concepts, with finetuning reducing the number of concepts and shifting them to more relevant categories.
Read Full Article
8 Likes
Arxiv
3d
160
Image Credit: Arxiv
Physics-informed neural networks for hidden boundary detection and flow field reconstruction
- A physics-informed neural network (PINN) framework is developed for detecting hidden solid boundaries and reconstructing flow fields from sparse observations in fluid mechanics.
- The PINN framework enforces no-slip/no-penetration boundary conditions and conservation laws of fluid dynamics while inferring the presence, shape, and motion of solid boundaries.
- The method successfully reconstructs flow fields and identifies solid boundaries using partial flow field data in various scenarios, including incompressible and compressible flows.
- The proposed method demonstrates robustness and versatility, making it suitable for applications with limited data availability.
Read Full Article
9 Likes
Arxiv
3d
280
Image Credit: Arxiv
Controlled Latent Diffusion Models for 3D Porous Media Reconstruction
- Three-dimensional digital reconstruction of porous media presents a fundamental challenge in geoscience.
- A computational framework, using latent diffusion models within the EDM framework, is introduced to address this challenge.
- The approach reduces dimensionality and enables the generation of larger volumes than previously possible with diffusion models.
- Extensive testing on four distinct rock types demonstrates that the framework achieves better generation quality and establishes a new state-of-the-art for digital rock physics applications.
Read Full Article
16 Likes
Arxiv
3d
0
Image Credit: Arxiv
New universal operator approximation theorem for encoder-decoder architectures (Preprint)
- A new universal operator approximation theorem has been presented for encoder-decoder architectures.
- The theorem focuses on approximating continuous operators in normed or metric spaces.
- The study explores the case where the approximating operator sequence can be chosen independently of the compact sets.
- The new approximation property tailored to encoder-decoder architectures ensures uniform convergence on every compact subset.
Read Full Article
Like
Arxiv
3d
48
Image Credit: Arxiv
Inductive Graph Representation Learning with Quantum Graph Neural Networks
- A versatile Quantum Graph Neural Network (QGNN) framework is proposed Integrating established techniques for inductive representation learning on graphs with parametrized quantum convolutional and pooling layers Benchmarked on a node regression task with the QM9 dataset, achieving performance comparable to classical Graph Neural Networks (GNNs) The quantum approach exhibits robust generalization across molecules with varying numbers of atoms, outperforming classical GNNs The QGNN framework demonstrates scalability without barren plateaus as the number of qubits increases
Read Full Article
2 Likes
For uninterrupted reading, download the app