Recursive neural programs: A differentiable framework for learning compositional part-whole hierarchies and image grammars.

This paper presents results on applying deep neural networks to analysis of images from borescope inspections of large turbofan engines, carried out in the field. Such inspections are done as a part of routine monitoring and maintenance as well as an initial, investigative response to alerts from automatic monitoring systems, pilots or engineers. Across GE’s commercial engines fleet, a substantial number of images have been gathered this way. The deep learning techniques that have come out of computer vision and machine learning research in the last decade offer new possibilities for analyzing and mining such data collections. This paper presents initial results on separating borescope images from images created with regular digital cameras, as well as classifying images containing various engine parts, with average accuracy of 95% and 77%, respectively, on unseen validation data.

Domain adaptation in image semantic segmentation has attracted more and more attention from computer vision and machine learning researchers. While the high cost of manual annotation is an unavoidable bottleneck in the semantic segmentation task, it is a high-quality solution to adopt pixel-level annotation from synthetic data, which provides additional support for deep learning training. Numerous studies have attempted to comprehensively investigate deep domain adaptation, but there is less focus on the sub-direction of the semantic segmentation task. This paper is devoted to this new topic in transfer learning. First, we describe the terminology and background concepts in this field. Next, the main datasets and evaluation metrics are introduced. Then, we classify the current research methods and introduce their contributions. Moreover, the quantitative results of the methods involved are compared, and the results are discussed. Finally, we suggest future research directions for this field. We believe researchers who are interested in this field will find this work to be an effective reference.

The informed sampler: A discriminative approach to Bayesian inference in generative computer vision models

A critical task for system operators is the precise identification of the root causes underlying an error situation. This identification is fundamental in deciding optimal maintenance actions, such as replacing a component versus calibrating it. However, the actual causes of an error are often neither measured nor unique. The measured quantities are the result of complex interactions between different error causes and system variables. Root cause identification in this context becomes a matter of inferring hidden causes from their measurable effects. This challenge is notably pronounced in cyber-physical systems comprising control loops. Control mechanisms, integral to maintaining system performance, introduce a layer of complexity in diagnostics and ultimately complicate the isolation of the underlying causes of errors. To address this challenge, we introduce a two-step approach to derive the hidden causes as a statistical inference task. First, we develop a generative model leveraging existing control software and expert-based insights into the mechanisms of errors, i.e., a simulator of synthetic data given some hidden error causes. Then, we transform the generative model into a probabilistic program on which statistical inference can be executed within a probabilistic programming language framework. This inference effectively estimates the hidden causes given some measured data from the system. Being intrinsically a statistical approach, these inferences come with a confidence interval. We applied this methodology to an industrial printer’s sheet transport belt, operating in a closed-loop configuration. Our approach successfully discerned the contributions of three distinct hidden causes to the belt’s deviation from its intended position. This paper highlights the efficacy of generative modeling followed by a probabilistic programming approach in unraveling complex interactions within cyber-physical systems for optimal maintenance.

Retrieval of flower videos based on a query with multiple species of flowers

Image classification serves as the cornerstone of computer vision, traditionally achieved through discriminative models based on deep neural networks. Recent advancements have introduced classification methods derived from generative models, which offer the advantage of zero-shot classification. However, these methods suffer from two main drawbacks: high computational overhead and inferior performance compared to discriminative models. Inspired by the coordinated cognitive processes of rapid-slow pathway interactions in the human brain during visual signal recognition, we propose the Diffusion-Based Discriminative Model Enhancement Framework (DBMEF). This framework seamlessly integrates discriminative and generative models in a training-free manner, leveraging discriminative models for initial predictions and endowing deep neural networks with rethinking capabilities via diffusion models. Consequently, DBMEF can effectively enhance the classification accuracy and generalization capability of discriminative models in a plug-and-play manner. We have conducted extensive experiments across 17 prevalent deep model architectures with different training methods, including both CNN-based models such as ResNet and Transformer-based models like ViT, to demonstrate the effectiveness of the proposed DBMEF.Specifically, the framework yields a 1.51% performance improvement for ResNet-50 on the ImageNet dataset and 3.02% on the ImageNet-A dataset. In conclusion, our research introduces a novel paradigm for image classification, demonstrating stable improvements across different datasets and neural networks.

Image inpainting, a pivotal technology for restoring damaged regions of images, has emerged as a significant research focus in computer vision. This review systematically surveys recent advances in deep learning-based image inpainting. We begin by categorizing prevailing methods into three groups based on their core architectures: Convolutional Neural Networks (CNNs), Generative Models, and Transformers. Through a comparative analysis of their symmetric versus asymmetric network architectures, applicable scenarios, and performance bottlenecks, we provide a critical discussion of the strengths and limitations inherent to each approach. The evolution of underlying design principles, such as symmetry, and the corresponding solutions to core challenges are also discussed. Furthermore, we introduce key benchmark datasets and commonly used image quality assessment metrics, offering a multidimensional framework for evaluation. We highlight that mainstream datasets collectively foster a greenhouse-like evaluation environment detached from real-world complexities and that existing metrics are critically misaligned with the fundamental objective of inpainting: generating plausible new content. Finally, we summarize the prevailing challenges in current deep learning-based inpainting research and outline promising future directions. We highlight critical issues, such as enhancing restoration quality, reducing computational costs, and broadening application scenarios, thereby providing valuable insights for subsequent research.

Reinforcement learning for neural architecture search: A review

Generative models that allow synthesis of realistic 3D models have been of interest in computer vision and graphics for over 2 decades. While traditional methods use morphable models for this task, more recent works have adopted powerful tools from the 2D image domain such as generative adversarial networks, neural fields and diffusion models, and have achieved impressive results. The question of which tools are most suitable for applications such as reconstructing 3D geometry from partial data, and creating digital 3D content remains open. This report documents the program and outcomes of Dagstuhl Seminar 25202 titled "Generative Models for 3D Vision". This meeting of 25 researchers covered a variety of topics such as generative models and priors for 2D tasks, medical applications, and digital representations of humans, including how to evaluate and benchmark different methods. We summarise the discussions, presentations, and results of this seminar.

Generative AI models represent a significant advancement in content creation capabilities but face substantial challenges when deployed at the network edge due to inherent resource constraints. This article examines comprehensive optimization strategies for enabling generative AI functionality on edge devices without requiring cloud connectivity. The exponential growth in model size has created a widening gap between computational requirements and the limited resources available in edge environments. Through systematic model compression, architectural redesign, and hardware-software co-optimization, generative models can achieve dramatic efficiency improvements while maintaining acceptable quality thresholds. The compression techniques examined include pruning methodologies that systematically eliminate redundant parameters, quantization approaches that reduce numerical precision, and knowledge distillation methods that transfer capabilities from larger models to compact alternatives. Architectural innovations such as modified attention mechanisms, conditional computation, and neural architecture search further enhance efficiency by fundamentally rethinking model design for resource-constrained environments. The integration of these techniques with hardware-specific optimizations and specialized software frameworks enables practical deployment across diverse application domains. Real-world implementations in speech processing, computer vision, and industrial IoT demonstrate that properly optimized generative models can operate within edge constraints while delivering near-real-time performance and maintaining high-quality outputs. These advancements empower industries to leverage generative AI capabilities in scenarios where privacy concerns, connectivity limitations, or latency requirements make cloud-based processing impractical.

Network neuroscience is the emerging discipline concerned with investigating the complex patterns of interconnections found in neural systems, and identifying principles with which to understand them. Within this discipline, one particularly powerful approach is network generative modelling, in which wiring rules are algorithmically implemented to produce synthetic network architectures with the same properties as observed in empirical network data. Successful models can highlight the principles by which a network is organized and potentially uncover the mechanisms by which it grows and develops. Here, we review the prospects and promise of generative models for network neuroscience. We begin with a primer on network generative models, with a discussion of compressibility and predictability, and utility in intuiting mechanisms, followed by a short history on their use in network science, broadly. We then discuss generative models in practice and application, paying particular attention to the critical need for cross-validation. Next, we review generative models of biological neural networks, both at the cellular and large-scale level, and across a variety of species including Caenorhabditis elegans, Drosophila, mouse, rat, cat, macaque and human. We offer a careful treatment of a few relevant distinctions, including differences between generative models and null models, sufficiency and redundancy, inferring and claiming mechanism, and functional and structural connectivity. We close with a discussion of future directions, outlining exciting frontiers both in empirical data collection efforts as well as in method and theory development that, together, further the utility of the generative network modelling approach for network neuroscience.

Artificial Intelligence in the American Healthcare Industry: Looking Forward to 2030

Connectome generative models, otherwise known as generative network models, provide insight into the wiring principles underpinning brain network organization. While these models can approximate numerous statistical properties of empirical networks, they typically fail to explicitly characterize an important contributor to brain organization-axonal growth. Emulating the chemoaffinity-guided axonal growth, we provide a novel generative model in which axons dynamically steer the direction of propagation based on distance-dependent chemoattractive forces acting on their growth cones. This simple dynamic growth mechanism, despite being solely geometry-dependent, is shown to generate axonal fiber bundles with brain-like geometry and features of complex network architecture consistent with the human brain, including lognormally distributed connectivity weights, scale-free nodal degrees, small-worldness, and modularity. We demonstrate that our model parameters can be fitted to individual connectomes, enabling connectome dimensionality reduction and comparison of parameters between groups. Our work offers an opportunity to bridge studies of axon guidance and connectome development, providing new avenues for understanding neural development from a computational perspective.

Computer Vision and Machine Learning are two branches of Artificial Intelligence which in combination have granted computational systems the ability of artificial vision and perception. Nowadays, human-machine interaction and the analysis of human behavior can be achieved using cameras. The subject of the current thesis is the implementation of a system for automatic recognition of static hand gestures in real-time videos with cameras, using modern techniques and algorithms of Machine Learning and Deep Learning. To begin with, we make an introduction into Computer Vision and its applications and we describe the problem of hand gesture recognition. Next, we refer to the field of Machine Learning and the main categories of the problems it includes. We describe the methodology of training a model and ways to measure its performance, focusing on Classification problems. Also, we introduce Deep Learning and Artificial Neural Networks. We present the most common network architectures used nowadays and the procedure of training a Neural Network and methods for optimization. Also, we mention the use of Convolutional Neural Networks for solving Computer Vision problems and we present some of the most known algorithms used for this task. Finally, we present our implementation of the hand gesture recognition tool, using Python programming language. We apply innovative algorithms of Convolutional Neural Networks for the detection of several key points of the hand and we train a Support Vector Machine and a Feedforward Neural Network model for the classification of those points into several predetermined static gestures. The implementation is targeted for practical real-time applications, runs using the CPU of the computer and requires a common RGB camera.

North in the head: spatial reference frame and map orientation