menu
techminis

A naukri.com initiative

google-web-stories
Home

>

ML News

ML News

source image

Arxiv

3h

read

116

img
dot

Image Credit: Arxiv

Probabilistic Trust Intervals for Out of Distribution Detection

  • Researchers propose a novel technique to enhance out-of-distribution (OOD) detection in pre-trained deep learning networks without altering their original parameters.
  • The approach defines probabilistic trust intervals for each network weight using in-distribution data and samples additional weight values during inference.
  • By quantifying the disagreements among outputs, the method achieves improved OOD detection performance compared to various baseline methods.
  • The proposed approach demonstrates robustness in identifying corrupted and adversarial inputs, without requiring OOD samples during training.

Read Full Article

like

7 Likes

source image

Arxiv

3h

read

210

img
dot

Image Credit: Arxiv

An Empirical Study: Extensive Deep Temporal Point Process

  • Temporal point process is commonly used to model asynchronous event sequences featuring occurrence timestamps.
  • Deep neural networks are emerging as a promising choice for capturing patterns in temporal point process.
  • This paper reviews recent research emphasis and difficulties in modeling asynchronous event sequences with deep temporal point process.
  • The study introduces recently proposed models and experiments to evaluate their performance.

Read Full Article

like

12 Likes

source image

Arxiv

3h

read

243

img
dot

Image Credit: Arxiv

Towards efficient representation identification in supervised learning

  • Humans have a remarkable ability to disentangle complex sensory inputs into simple factors of variation without much supervision.
  • Existing works on disentanglement require a lot of auxiliary information, which may not provide conditional independence over the factors of variation.
  • This research explores how to achieve disentanglement when the auxiliary information does not ensure conditional independence.
  • The study shows theoretically and experimentally that disentanglement is possible with reduced auxiliary information dimension.

Read Full Article

like

14 Likes

source image

Arxiv

3h

read

70

img
dot

Image Credit: Arxiv

Level Up with ML Vulnerability Identification: Leveraging Domain Constraints in Feature Space for Robust Android Malware Detection

  • Machine Learning (ML) promises to enhance the efficacy of Android Malware Detection (AMD) but is vulnerable to realistic evasion attacks.
  • Defenders aim to identify susceptible regions in the feature space where ML models are prone to deception.
  • A proposed approach introduces a new interpretation of Android domain constraints in the feature space and employs a novel technique to learn them.
  • Empirical evaluations show effective detection of Adversarial Examples (AEs) using learned domain constraints and improved robustness against realizable AEs generated by unknown problem-space transformations.

Read Full Article

like

4 Likes

source image

Arxiv

3h

read

23

img
dot

Image Credit: Arxiv

The Numerical Stability of Hyperbolic Representation Learning

  • The hyperbolic space is capable of embedding trees with arbitrarily small distortion, making it suitable for representing hierarchical datasets.
  • However, training hyperbolic learning models can lead to numerical instability and NaN problems due to the exponential growth property of the hyperbolic space.
  • A study compares two popular models, the Poincaré ball and the Lorentz model, and finds that the Lorentz model has superior numerical stability and optimization performance.
  • Additionally, an Euclidean parametrization of the hyperbolic space is proposed to alleviate numerical limitations, which also improves the performance of hyperbolic SVM.

Read Full Article

like

1 Like

source image

Arxiv

3h

read

70

img
dot

Image Credit: Arxiv

Zero-Shot Conditioning of Score-Based Diffusion Models by Neuro-Symbolic Constraints

  • Score-based diffusion models have emerged as effective approaches for both conditional and unconditional generation.
  • A new method is proposed to sample from the conditional distribution under arbitrary logical constraints without additional training.
  • The method manipulates the learned score to sample from an un-normalized distribution based on user-defined constraints.
  • The approach shows effectiveness in approximating conditional distributions for tabular data, images, and time series.

Read Full Article

like

4 Likes

source image

Arxiv

3h

read

103

img
dot

Image Credit: Arxiv

Tackling Intertwined Data and Device Heterogeneities in Federated Learning with Unlimited Staleness

  • Federated Learning (FL) can be affected by data and device heterogeneities, caused by clients' different local data distributions and latencies in uploading model updates.
  • Traditional FL schemes consider these heterogeneities as separate aspects, but in practical scenarios, they are intertwined.
  • A new FL framework is presented in this paper, which converts stale model updates into unstale ones to tackle intertwined heterogeneities.
  • The approach estimates the distributions of clients' local training data from stale model updates and improves model accuracy by up to 25%.

Read Full Article

like

6 Likes

source image

Arxiv

3h

read

63

img
dot

Image Credit: Arxiv

Perfect Alignment May be Poisonous to Graph Contrastive Learning

  • Graph Contrastive Learning (GCL) focuses on aligning positive pairs and separating negative ones to learn node representations.
  • This paper addresses the connection between augmentation and downstream performance in GCL.
  • Findings reveal that GCL mainly contributes to downstream tasks by separating different classes.
  • Perfect alignment and augmentation overlap may not lead to the best downstream performance, so specifically designed augmentations are needed.

Read Full Article

like

3 Likes

source image

Arxiv

3h

read

200

img
dot

Image Credit: Arxiv

Enhancing the Performance of Neural Networks Through Causal Discovery and Integration of Domain Knowledge

  • Researchers have developed a methodology called causality-informed neural network (CINN) to improve the predictive performance of neural networks.
  • CINN leverages three steps to encode hierarchical causality structure into the neural network, discovered through causal discovery from observational data.
  • The discovered causal relationships are systematically encoded into the neural network's architecture and loss function, preserving the relative order and co-learning of different types of nodes.
  • Computational experiments show that CINN outperforms other state-of-the-art methods in predictive performance, highlighting the value of integrating causal knowledge.

Read Full Article

like

12 Likes

source image

Arxiv

3h

read

50

img
dot

Image Credit: Arxiv

Prompted Contextual Vectors for Spear-Phishing Detection

  • Spear-phishing attacks present a significant security challenge.
  • A detection approach based on novel document vectorization method is proposed.
  • The method achieves a 91% F1 score in identifying LLM-generated spear-phishing emails.
  • Key contributions include a novel document vectorization method and a publicly available dataset of high-quality spear-phishing emails.

Read Full Article

like

3 Likes

source image

Arxiv

3h

read

10

img
dot

Image Credit: Arxiv

Tracking Changing Probabilities via Dynamic Learners

  • Consider a predictor, a learner, whose input is a stream of discrete items.
  • The predictor's task is probabilistic multiclass prediction, predicting the next item with probabilities.
  • The predictor keeps track of item proportions and adjusts probabilities based on revealed items.
  • Various predictors, such as sparse moving averages, are designed and evaluated for the task.

Read Full Article

like

Like

source image

Arxiv

3h

read

197

img
dot

Image Credit: Arxiv

Advancing Additive Manufacturing through Deep Learning: A Comprehensive Review of Current Progress and Future Challenges

  • This paper presents the first comprehensive literature review of deep learning (DL) applications in additive manufacturing (AM).
  • The review covers three major areas of AM: design for AM, AM modeling, and monitoring and control in AM.
  • The analysis reveals a trend towards using deep generative models for generative design in AM, and incorporating process physics into DL models to improve AM process modeling.
  • The paper summarizes the current challenges and recommends areas for further investigation, including generalizing DL models for a wide range of geometry types and incorporating deep generative models to address limited and noisy AM data.

Read Full Article

like

11 Likes

source image

Arxiv

3h

read

63

img
dot

Image Credit: Arxiv

Locally Convex Global Loss Network for Decision-Focused Learning

  • In decision-making problems under uncertainty, predicting unknown parameters is often considered independent of the optimization part.
  • Decision-focused learning (DFL) is a task-oriented framework that integrates prediction and optimization by adapting the predictive model to give better decisions for the corresponding task.
  • In this paper, the authors propose Locally Convex Global Loss Network (LCGLN), a global surrogate loss model that can be implemented in a general DFL paradigm.
  • LCGLN learns task loss via a partial input convex neural network which is guaranteed to be convex for chosen inputs while keeping the non-convex global structure for the other inputs.

Read Full Article

like

3 Likes

source image

Arxiv

3h

read

63

img
dot

Image Credit: Arxiv

The Effectiveness of Local Updates for Decentralized Learning under Data Heterogeneity

  • We revisit two fundamental decentralized optimization methods, Decentralized Gradient Tracking (DGT) and Decentralized Gradient Descent (DGD), with multiple local updates.
  • Incorporating local update steps can reduce communication complexity for strongly convex and smooth loss functions.
  • Increasing the number of additional local updates can effectively reduce communication costs when data heterogeneity is low and the network is well-connected.
  • Employing local updates in DGD achieves exact linear convergence under the Polyak-Łojasiewicz (PL) condition.

Read Full Article

like

3 Likes

source image

Arxiv

3h

read

260

img
dot

Image Credit: Arxiv

Distance-Restricted Explanations: Theoretical Underpinnings & Efficient Implementation

  • Explainable artificial intelligence (XAI) aims to help human decision-makers understand complex machine learning models.
  • Logic-based XAI offers rigorous explanations but can be complex, especially for highly complex ML models.
  • Recent work proposes distance-restricted explanations that are rigorous within a small input distance.
  • This paper investigates algorithms to scale up logic-based explainers for computing explanations with a large number of inputs.

Read Full Article

like

15 Likes

For uninterrupted reading, download the app