A naukri.com initiative
ML News
Arxiv
4d
348
Image Credit: Arxiv
Faithful and Accurate Self-Attention Attribution for Message Passing Neural Networks via the Computation Tree Viewpoint
- Researchers propose GATT, a method for calculating edge attributions in self-attention message passing neural networks (MPNNs) based on the computation tree.
- GATT aims to bridge the gap between the widespread usage of attention-based MPNNs (Att-GNNs) and their potential explainability.
- The proposed method improves edge attribution scores, demonstrating effectiveness in model explanation, faithfulness, explanation accuracy, and case studies.
- The code for GATT is available on GitHub for reference.
Read Full Article
20 Likes
Arxiv
4d
48
Image Credit: Arxiv
Opening the Black Box: predicting the trainability of deep neural networks with reconstruction entropy
- An important challenge in machine learning is to predict the initial conditions under which a given neural network will be trainable.
- A method for predicting the trainable regime in parameter space for deep feedforward neural networks (DNNs) is presented.
- The method involves reconstructing the input from subsequent activation layers via a cascade of single-layer auxiliary networks.
- The method shows promise in reducing overall training time and generalizes to residual neural networks (ResNets) and convolutional neural networks (CNNs).
Read Full Article
2 Likes
Arxiv
4d
348
Image Credit: Arxiv
Spectral Self-supervised Feature Selection
- Choosing a meaningful subset of features from high-dimensional observations in unsupervised settings can greatly enhance the accuracy of downstream analysis.
- The proposed method is a self-supervised graph-based approach for unsupervised feature selection.
- It involves computing robust pseudo-labels using the graph Laplacian's eigenvectors and a model stability criterion.
- Experiments on real-world datasets demonstrate the method's effectiveness, especially in biological datasets.
Read Full Article
20 Likes
Arxiv
4d
68
Image Credit: Arxiv
Lifelong Graph Learning for Graph Summarization
- Researchers investigate lifelong graph learning for graph summarization.
- Neural networks are used to summarize web graphs, considering the heterogeneity and temporal changes.
- Graph-MLP and GraphSAINT, along with an MLP baseline, are used for summarization.
- Experiments show that 1-hop information is predominantly used, even in 2-hop summarization, with a drop in accuracy over a ten-year time period.
Read Full Article
4 Likes
Arxiv
4d
388
Image Credit: Arxiv
Error Bounds For Gaussian Process Regression Under Bounded Support Noise With Applications To Safety Certification
- Gaussian Process Regression (GPR) is a powerful method for learning complex functions from noisy data.
- This paper introduces novel error bounds for GPR under bounded support noise.
- The derived probabilistic and deterministic bounds are tighter than existing state-of-the-art bounds.
- The bounds can be combined with stochastic barrier functions to quantify the safety probability of unknown dynamical systems.
Read Full Article
23 Likes
Arxiv
4d
372
Image Credit: Arxiv
Assessing Modality Bias in Video Question Answering Benchmarks with Multimodal Large Language Models
- Multimodal large language models (MLLMs) can process visual, textual, and auditory data.
- Existing video question-answering benchmarks often exhibit bias towards a single modality.
- The modality importance score (MIS) is introduced to identify and assess modality bias.
- MLLM-derived MIS can guide the curation of modality-balanced datasets to enhance multimodal learning.
Read Full Article
22 Likes
Arxiv
4d
206
Image Credit: Arxiv
MicroFlow: An Efficient Rust-Based Inference Engine for TinyML
- MicroFlow is an open-source TinyML framework for the deployment of Neural Networks (NNs) on embedded systems.
- It is built using the Rust programming language and is suitable for resource-constrained devices.
- MicroFlow enables the successful deployment of NNs on devices with limited memory and processing power.
- It achieves efficient inference with less Flash and RAM memory usage compared to other solutions.
Read Full Article
12 Likes
Arxiv
4d
320
Image Credit: Arxiv
Replacing Paths with Connection-Biased Attention for Knowledge Graph Completion
- Knowledge graph completion aims to identify additional facts in the KG.
- Recent developments in this field have explored KG completion in the inductive setting.
- The CBLiP model introduces connection-biased attention and entity role embeddings to improve performance.
- CBLiP outperforms models that do not use path information and is faster than models that utilize path information.
Read Full Article
19 Likes
Arxiv
4d
392
Image Credit: Arxiv
Differentially Private Release and Learning of Threshold Functions
- The research paper introduces new upper and lower bounds on the sample complexity of differentially private algorithms for releasing approximate answers to threshold functions.
- The paper proves a nontrivial lower bound for releasing thresholds with differential privacy, showing that it is impossible over an infinite domain and requires a sample complexity that grows with the size of the domain.
- The paper also presents an algorithm for releasing thresholds with improved sample complexity compared to previous bounds.
- The research results have implications for learning distributions, properly PAC learning thresholds with differential privacy, and provide separation between learning with and without privacy.
Read Full Article
23 Likes
Arxiv
4d
202
Image Credit: Arxiv
Learning Low Degree Hypergraphs
- We study the problem of learning a hypergraph via edge detecting queries.
- Learning a hypergraph with m edges of max size d requires Ω((2m/d)^(d/2)) queries.
- We identify hypermatchings and low-degree near-uniform hypergraphs as learnable with poly(n) queries.
- For hypergraphs with max degree Δ and edge size ratio ρ, we give a non-adaptive algorithm with O((2n)^(ρΔ+1)log^2n) queries.
Read Full Article
12 Likes
Arxiv
4d
0
Image Credit: Arxiv
Variational measurement-based quantum computation for generative modeling
- Measurement-based quantum computation (MBQC) offers a unique paradigm for designing quantum algorithms.
- MBQC algorithms can embrace the inherent randomness of quantum measurements and treat them as a resource for computation.
- The proposed variational MBQC algorithm allows adjusting the degree of randomness in the computation.
- This additional randomness can lead to gains in expressivity and learning performance for generative modeling tasks.
Read Full Article
Like
Arxiv
4d
8
Image Credit: Arxiv
A Hybrid Probabilistic Battery Health Management Approach for Robust Inspection Drone Operations
- Researchers have developed a hybrid probabilistic approach for battery health management in inspection drones.
- The approach combines physics-based discharge and probabilistic error-correction models to predict battery end-of-discharge voltage.
- The methodology was tested on a dataset obtained from real inspection drones used for wind turbine inspections.
- The hybrid approach demonstrated a 14.8% improvement in probabilistic accuracy compared to the best probabilistic method.
Read Full Article
Like
Arxiv
4d
380
Image Credit: Arxiv
LLM Processes: Numerical Predictive Distributions Conditioned on Natural Language
- Machine learning practitioners face challenges in integrating prior knowledge into predictive models.
- The goal is to build a regression model that can make probabilistic predictions guided by natural language text.
- Large Language Models (LLMs) provide an interface to incorporate expert insights in natural language.
- LLM Processes explore strategies for eliciting coherent numerical predictive distributions from LLMs.
Read Full Article
22 Likes
Arxiv
4d
190
Image Credit: Arxiv
Local Causal Discovery for Structural Evidence of Direct Discrimination
- A new method called local discovery for direct discrimination (LD3) has been introduced to uncover structural evidence of direct unfairness by identifying the causal parents of an outcome variable.
- LD3 performs a linear number of conditional independence tests relative to variable set size and allows for latent confounding.
- LD3 returns a valid adjustment set (VAS) under a new graphical criterion for the weighted controlled direct effect, enabling assessment of direct discrimination.
- LD3 shows more time efficiency and provides more plausible results compared to baselines in the analysis of causal fairness in complex decision systems.
Read Full Article
11 Likes
Arxiv
4d
320
Image Credit: Arxiv
Factor Augmented Tensor-on-Tensor Neural Networks
- This paper introduces Factor Augmented Tensor-on-Tensor Neural Networks (FATTNN) for tensor-on-tensor regression.
- FATTNN integrates tensor factor models into deep neural networks to handle nonlinearity between complex data structures.
- The proposed methods offer improved prediction accuracy and computational efficiency compared to traditional statistical models and conventional deep learning approaches.
- Empirical results from simulation studies and real-world applications show the superiority of FATTNN over benchmark methods.
Read Full Article
19 Likes
For uninterrupted reading, download the app