Refine
Document Type
- conference proceeding (article) (6)
- Preprint (5)
- Part of Periodical (2)
- Report (1)
Is part of the Bibliography
- no (14)
Keywords
- Convergence (3)
- Adam optimizer (2)
- Konvergenz 〈Informationstechnik〉 (2)
- Maschinelles Lernen (2)
- Neural networks (2)
- Neuronales Netz (2)
- Optimierungsalgorithmus (2)
- kinematic singularities (2)
- ADAM Optimizer (1)
- CNN (1)
Institute
- Fakultät Informatik und Mathematik (12)
- Research Center for Artificial Intelligence - RCAI (7)
- Research Center of Energy and Resources - RCER (3)
- Hochschulleitung/Hochschulverwaltung (2)
- Research Center of Biomedical Engineering - RCBE (2)
- Zentrum für Forschung und Transfer (ZFT ab 2024; vorher: IAFW) (2)
- Research Center of Health Sciences and Technology - RCHST (1)
Begutachtungsstatus
- peer-reviewed (3)
Forschung 2019
(2019)
Adaptive Moment Estimation (ADAM) is a very popular training algorithm for deep neural networks and belongs to the family of adaptive gradient descent optimizers. However to the best of the authors knowledge no complete convergence analysis exists for ADAM. The contribution of this paper is a method for the local convergence analysis in batch mode for a deterministic fixed training set, which gives necessary conditions for the hyperparameters of the ADAM algorithm. Due to the local nature of the arguments the objective function can be non-convex but must be at least twice continuously differentiable. Then we apply this procedure to other adaptive gradient descent algorithms and show for most of them local convergence with hyperparameter bounds.
Adaptive Moment Estimation (Adam) is a very popular training algorithm for deep neural networks, implemented in many machine learning frameworks. To the best of the authors knowledge no complete convergence analysis exists for Adam. The contribution of this paper is a method for the local convergence analysis in batch mode for a deterministic fixed training set, which gives necessary conditions for the hyperparameters of the Adam algorithm. Due to the local nature of the arguments the objective function can be non-convex but must be at least twice continuously differentiable.
The main issues in many image processing applications are
object recognition and detection of objects, which answers the questions whether an object is present and if it is present, where it is located. Popular object detection algorithms like YOLO use a regression formulation for the whole problem, especially for the bounding box parameters. In production industry the setting usually is different: One usually knows the object type and rather wants to know with high precision where the object is. We study a prototype application in this area where we identify the rotation of an object in a plane. To solve this problem use a regression approach with a CNN architecture as a function approximator. We compare our results to standard image processing algorithms, which do not use neural networks, and present quantitative results on the accuracy.
CNNs seem at least competitive to classical image processing.
One of the most popular training algorithms for deep neural networks is the Adaptive Moment Estimation (Adam) introduced by Kingma and Ba. Despite its success in many applications there is no satisfactory convergence analysis: only local convergence can be shown for batch mode under some restrictions on the hyperparameters, counterexamples exist for incremental mode. Recent results show that for simple quadratic objective functions limit cycles of period 2 exist in batch mode, but only for atypical hyperparameters, and only for the algorithm without bias correction. We extend the convergence analysis to all choices of the hyperparameters for quadratic functions. This finally answers the question of convergence for Adam in batch mode to the negative. We analyze the stability of these limit cycles and relate our analysis to other results where approximate convergence was shown, but under the additional assumption of bounded gradients which does not apply to quadratic functions. The investigation heavily relies on the use of computer algebra due to the complexity of the equations.
A basic task in the design of an industrial robot application
is the relative placement of robot and workpiece. Process points are defined in Cartesian coordinates relative to the workpiece coordinate system, and the workpiece has to be located such that the robot can reach all points. Finding such a location is still an iterative procedure based on the developers' intuition. One difficulty is the choice of one of the several solutions of the backward transform of a typical 6R robot.
We present a novel algorithm that simultaneously optimizes the workpiece location and the robot configuration at all process points using higher order optimization algorithms. A key ingredient is the extension of the robot with a virtual prismatic axis. The practical feasibility of the approach is shown with an example using a commercial industrial robot.
Die Schätzung der 6D-Pose bekannter Objekte findet Anwendungen in der Robotik, der Luft- und Raumfahrtsteuerung und automatisierten Produktionsumgebungen. Es ist nach wie vor gängige Praxis, klassische Bildverarbeitungsmethoden zu verwenden, um eine hohe Präzision zu erreichen. Diese Methoden erfordern jedoch eine manuelle Parametrisierung der verwendeten Erkennungswerkzeuge für jedes einzelne Objekt.
Dies wirft die Frage auf, ob maschinelle Lerntechniken, insb. convolutional neural networks, Netzwerke, so trainiert werden können, dass sie die in der Industrie geforderte Präzision ohne individuelle Programmierung erreichen können.
Darüber sollen in Produktionsumgebungen Umwelteinflüsse wie Lichtverhältnisse minimiert werden.
Es werden verschiedene Netzstrukturen entwickelt und auf ihre Leistung bei diesem Szenario untersucht, mit 1 bis 6 Freiheitsgraden in der Aufgabenstellung und einer oder mehreren Kameras. Zusätzlich wird ein Ansatz mit reinforcement learning entwickelt und untersucht.
We present a new method for studying the workspace of 6R robots using a four-dimensional representation. This is applied to numerically determine connected workspace components with a constant number of solutions of the inverse kinematics, and to describe critical values that form the boundaries these regions. To define and maximize uniqueness domains, nonsingular transition of these boundaries are investigated. This provides a method to determine workspace regions in which the robot can move avoiding configuration changes and singularities.
We present a theoretical and numerical analysis of the kinematics for the Transpressor, a cuspidal 6R robot. It admits up to 16 inverse kinematics solutions which are described geometrically. For special target poses, we provide the solutions analytically and present a simple numerical solver for the general case. Moreover, an analytical estimate of the Jacobian determinant on a path between two solutions proves cuspidality for a class of robots similar to the transpressor.
Object detection and recognition is a core task in machine vision. Both the position and the rotation of the objects are of interest. Common algorithms for object recognition based on neural networks use a regression formulation and are aimed at changing environments. In industrial applications, the setting is different: One would like to determine with high accuracy, the pose of known objects. We are investigating a prototype application in this field, where we identify the rotation of an object in a plane. We use common network architectures and provide an overview of how well they are suited for this purpose. In addition, we compare our results with standard image processing algorithms, paying particular attention to the accuracy with metrics relevant in the industry. Moreover, an U-shape phenomenon with disturbed data occurred during the experiments, which is partially explained analytically.
