Publication Server of Albstadt-Sigmaringen University
Not a member yet
    662 research outputs found

    Asynchronous Derivative-Free Learning Solving Synaptic Credit Assignment in Recurrent Neural Networks

    Get PDF
    In the quest for biologically faithful, derivative-free algorithms, the primary aim is to mimic the efficiency and adaptability observed in neural systems, while also attaining performance comparable to gradient-based methods. However, traditional gradient-based backpropagation (BP) algorithms like Back-Propagation Through Time (BPTT) face criticism for relying on i) synaptic credit assignment that relies on non-local synaptic information, as well as issues such as the ii) update locking and iii) weight symmetry problems. To address these concerns, we introduce an optimizer, Dopamine, for Weight Perturbation (WP) learning—a reward-based learning algorithm. Dopamine mitigates the aforementioned issues by i) leveraging a global learning signal to acquire "local error information" while adhering to locality constraints, ii) being gradient-free; it updates the parameters asynchronously during the backward pass, eliminating the need to freeze weights, and iii) ensuring that feedforward and feedback weights are not identical by perturbing the feedforward weights to get reward signal. We conducted a comparative evaluation of our Dopamine powered WP learning against the standalone WP, Stochastic Gradient Descent (SGD) and Adam Optimizers. The tests were conducted on synthetic time series generated by solving the Rossler and Lorenz equations and also on classical "Figure 8" task for time series prediction and system modeling. The results showcase the accelerated convergence of Dopamine model compared to the models trained with SGD and standard WP models. However, Adam exhibits a slight advantage inconvergence speed. Moreover, Dopamine’s parameter update operations are parallelizable and independent of time T, resulting in a total computational complexity of O(2NT+N) given a sequence of length T for N neural units. In contrast, the computational complexity of BPTT is O(NT+TN^2) due to the process of unrolling the network over time. Additionally, by bypassing the activation function, this algorithm provides an alternative to surrogate gradient descent in the training of spiking neural networks

    Quantification of Autophagosomes in Human Fibroblasts Using Cyto-ID® Staining and Cytation Imaging

    Get PDF
    As an essential process for the maintenance of cellular homeostasis and function, autophagy is responsible for the lysosome-mediated degradation of damaged proteins and organelles; therefore, dysregulation of autophagy in humans can lead to a variety of diseases. The link between impaired autophagy and disease highlights the need to investigate possible interventions to address dysregulations. One possible intervention is hyperthermia, which is described in this protocol. To investigate these interventions, a method for absolute quantification of autophagosomal compartments is required that allows comparison of autophagosomal activity under different conditions. Existing methods such as western blotting and immunohistochemistry for analysing the location and relative abundance of intracellular proteins associated with autophagy, or transmission electron microscopy (TEM), which are either very time-consuming, expensive, or both, are less suitable for this purpose. The method described in this protocol allows the absolute quantification of autophagosomes per cell in human fibroblasts using the CYTO-ID® Autophagy Detection Kit after heat therapy compared to a control. The Cyto-ID® assay is based on the use of a specific dye that selectively stains autophagic compartments, combined with an additional Hoechst 33342 dye for nuclear staining. The subsequent recognition of these stained compartments by the Cytation Imager enables the software to determine the number of autophagosomes per nucleus in living cells. Additionally, this absolute quantification uses an image-based method, and the protocol is easy to use and not time-consuming. Furthermore, the method is not only suitable for heat therapy but can also be adapted to any other desired therapy or substance

    KI-Verordnung und Datenschutz: Herausforderungen und Chancen

    No full text
    Die rasante Entwicklung von Künstlicher Intelligenz (KI) eröffnet zahlreiche Möglichkeiten und Vorteile in verschiedenen Bereichen. Gleichzeitig stellt sie jedoch erhebliche Herausforderungen im Bereich des Datenschutzes dar. Dieser Beitrag beleuchtet das Spannungsfeld zwischen KI und Datenschutz, indem er die wesentlichen rechtlichen Rahmenbedingungen, Arten von KI-Systemen und die damit verbundenen Pflichten und Verantwortlichkeiten erörtert. Mit der Einführung der Datenschutz-Grundverordnung (DS-GVO) im Mai 2018 wurde der Schutz personenbezogener Daten in Europa auf eine neue Grundlage gestellt. Parallel dazu hat sich die Nutzung von KI-Systemen weiterentwickelt, was neue Herausforderungen für den Datenschutz mit sich bringt. Die Hambacher Erklärung (2019) und die darauf basierenden Positionspapiere sowie die Orientierungshilfe der DSK zu KI (2024) haben diese Entwicklungen weiter geprägt. Die KI-Verordnung (KI-VO) der EU, die im Mai 2024 in Kraft trat, setzt nun neue Maßstäbe für den Umgang mit KI-Systemen

    Fettoptimierung in feinen Backwaren: Potentialanalyse für die Gemeinschaftsgastronomie und den privaten Haushalt

    Get PDF
    Feine Backwaren werden als Zwischenmahlzeit und als Dessert regelmäßig konsumiert, weisen jedoch meist einen hohen Zucker- und Fettgehalt sowie ein ungünstiges Fettsäuremuster auf. Anhand eines Sandkuchens werden Optionen zur Optimierung der Fettqualität durch (Teil-) Substitution durch Fette mit günstigerem Fettsäuremuster zum Referenzbackfett Butter sowie mit vollwertigen Lebensmitteln ermittelt. Eine (Teil-)substitution mit Rapsöl und Apfelmark reduziert den Anteil gesättigter Fettsäuren um ca. 80 % auf ~ 3 g/100 g, mit Magerquark um ca. 70 % auf 5 g/100 g, bei technologisch und sensorischer Machbarkeit. Nährwertoptimierte Rezepturalternativen sind sowohl in der Außer-Haus-Verpflegung als auch im privaten Haushalt als Standard zu implementieren

    Inertial Navigation – Importance of Initial Heading: A Constrained GNSS Receiver Approach

    Get PDF
    This paper describes scenarios for automated driving which are highly relevant on accurate initial heading estimation.We show thoughtfully designed experiments which are closely aligned with real world demands to explain the relevance and influence of initial heading estimation. We conclude that purely inertial navigation is highly dependent on accurate initial heading. On the basis of these experiments we show a constrained dual GNSS receiver approach which takes advantage of short baselines and known receiver baseline length to estimate initial heading without relying on the Earth’s magnetic field

    Unternehmensweite Datenstrategien in der Pharmabranche: Strategisches Erfolgspotenzial für eine datengetriebene Kundenzentrierung im digitalen Zeitalter

    Get PDF
    Wissenschaftliches E-Book "Unternehmensweite Datenstrategien in der Pharmabranche: Strategisches Erfolgspotenzial für eine datengetriebene Kundenzentrierung im digitalen Zeitalter", das die Entwicklung einer konzernweiten Datenstrategie für den TEVA-Ratiopharm-Konzern darstell

    224

    full texts

    662

    metadata records
    Updated in last 30 days.
    Publication Server of Albstadt-Sigmaringen University
    Access Repository Dashboard
    Do you manage Open Research Online? Become a CORE Member to access insider analytics, issue reports and manage access to outputs from your repository in the CORE Repository Dashboard! 👇