News zum MScTI
Aktuelles: Schließung der Vertiefung 'Biorobotics'
Leider werden die Professoren L. Masia und D. Häufle das ZITI zum kommenden Wintersemester verlassen. Ihre Forschung im Bereich 'Biorobotics' wird daher nicht mehr stattfinden und die Lehre in der zugehörigen Vertiefungsrichtung kann nicht mehr zuverlässig geleistet werden. Die Vertiefung 'Biorobotics' im MScTI muss daher leider ab sofort geschlossen werden.
Für unsere aktuellen Studierenden in diesem Bereich suchen wir nach Lösungen, wie sie ihr Studium so gut wie möglich zu Ende bringen können. Neue Bewerberinnen und Bewerber, die speziell an dieser Vertiefung interessiert waren, müssen nach Alternativen suchen.
ZITI Kolloquium
When & Where: Monday, 29.7., 4:15 pm, room U014 of OMZ (INF 350, floor -1)
Title: Efficient Hardware for Neural Networks
Speaker: Prof. Dr. Grace Li Zhang (TU Darmstadt)
Abstract: The last decade has witnessed significant breakthroughs of deep neural networks (DNNs) in many fields. These breakthroughs have been achieved at extremely high computation and memory cost. Accordingly, the increasing complexity of DNNs has led to a quest for efficient hardware platforms. In this talk, class-aware pruning is first presented to reduce the number of multiply-and-accumulate (MAC) operations in DNNs. Class-exclusion early-exit is then examined to reveal the target class before the last layer is reached. To accelerate DNNs, digital accelerators such as systolic array from Google can be used. Such an accelerator is composed of an array of processing elements to efficiently execute MAC operations in parallel. However, such accelerators suffer from high energy consumption. To reduce energy consumption of MAC operations, we select quantized weight values with good power and timing characteristics. To reduce energy consumption incurred by data movement, logic design of neural networks is presented. In the end, on-going research topics and future research plan will be summarized.
Bio: Grace Li Zhang received the Dr.-Ing. degree from the Technical University of Munich (TUM), in 2018. She joined TU Darmstadt in 2022 as a Tenure Track Assistant Professor. She leads the Hardware for Artificial Intelligence Group. Her research focuses on efficient hardware acceleration for AI algorithms and systems, AI computing with emerging devices, e.g., RRAM and optical components, circuit and system design methodologies for AI, explainability of AI and neuromorphic computing.
As usual, there will be snacks and drinks after the talk.
Anstehende Master Kolloquien
4.9.2024 - Hadi Ghaeni (10:00, online -
https://eu02web.zoom-x.de/j/67231519711?pwd=x4vkEVwUItxpVTS7gYzbewiRkAL…
): Analysis of the Suitability of Federated Learning Approaches for Quality Data of Eroding Products The transfer and storage of data is a controversial issue in modern industry. Operation Data from one or more machines or factory floors, such as the lubricating oil ageing data in this thesis, is classified as a trade secret. Companies must therefore ensure the security of this data. For this reason, data security is a very important consideration. Lubricating oil ageing data varies in many ways from machine to machine and is distributed across different sources. In order to train a model with classic Machine Learning (ML) methods based on this ageing data to determine the condition of the lubricating oil as a regression task, all of this machine data must first be collected at a central node, such as a data centre. The algorithm can then be trained on the collected data to solve the task. Federated learning (FL) is a method where the model is trained with decentralised data. In FL, each data source has a local model that is fed with the raw data from that source. The global model is trained with the computed local model parameters. These local parameters are aggregated into the global model. Depending on the application, there are different algorithms for aggregation. This thesis investigates the behaviour of the models in a classical ML and a FL en- vironment and additionally focuses on different aggregation algorithms for FL and their influence on the model quality. The model chosen for this work is a dense neural network with two hidden layers, which is trained in both a classical ML and a FL environment with real ageing data of industrial lubricating oils. The task is to predict the state of the lubricating oil. FedAvg, FedAvgm, FedMedian and QFedAvg are the aggregation algorithms analysed in this work. The centrally trained model achieves a coefficient of determination of 0.9 on the test set, while the federated server model achieves a coefficient of determination of 0.79. The investigation of a FL environment with real time series data shows accurate results for real time condition monitoring while preserving data privacy, and provides a reliable basis for predicting the remaining lifetime of lubricating oils.Fördermöglichkeiten für Studierende
Das ZITI bietet interessierten Studierenden finanzielle Unterstützung z.B. für Reisen zu Wettbewerben oder für eigene kleinere Projekte an. Weitere Infos finden Sie auf der verlinkten Seite.
Ältere News
- Die Einführungsveranstaltung für das Sommersemester 2024 fand am 15.4.2024 statt. Der neu angeschaffte Grill hat bei der anschließenden 'Thesis Fair' trotz des starken Windes gut funktioniert! Danke an alle Helferinnen und Helfer!
- Die Einführungsveranstaltung für unsere neuen Studierende des MScTI für das Wintersemester 2023/2024 fand am 16.10.2023 statt. Im Anschluss haben sich die Arbeitsgruppen mit Postern und Infos vorgestellt und auch Themen für Masterarbeiten präsentiert. Dazu waren insbesondere unsere aktuellen Studierenden herzlich eingeladen. Bei kühlem, aber klarem Wetter gab es Getränke und Grillgut. So konnten die neuen Studierenden mit den aktuellen Studierenden ins Gespräch kommen!
- Die Einführungsveranstaltung für neue Studierende des MScTI für das Sommersemester 2023 fand am 17.04.2023 statt.
Vergangene Master Kolloquien
- 31.7.2024 - Jonathan Klamroth : Design and characterization of two low-power memories and design of a leakage current compensation circuit for the readout ASIC of the XIDyn X-ray pixel detector
- 30.7.2024 - Arne Hegel: FBG-basierter optischer Kraftsensor am distalen Ende eines Glasfaser-Führungsdrahtes zur Verbesserung ferngesteuerter Katheterisierungssysteme
- 12.4.2024 - Gennadiy Knis: Automatic Generation of Physical Constrains for Implementing FPGAs and Their Optimization
- 2.2.2024 - Matthias Musch: Exergame using a Hybrid Control Strategy merging Exosuit and Functional Electrical Stimulation
- 21.12.2023 - Adil Omari: Entwurf eines Digital- Analog-Wandlers unter Verwendung von Stromdivision
- 15.12.2023 - Matthias Musch: Implementation and Optimization of a Modern FPGA Fabric with Open Source Tools
- 12.12.2023 - Ibrahima Kouruma: Design and development of a portable sensory system for biomechanical evaluation and motion-intention detection for lower-limb exoskeletons
- 5.12.2023 - Lisa Kuhn: Scalability of Bayesian Neural Network Inference Methods for Real-World Tasks
- 31.10.2023 - Tristan Strahler: Design of Electronic Interfaces and Test Infrastructure for a Novel Integrated Photonics Tensor Core
- 18.10.2023 - Eric Kern: Optimized Calibration for Analog Computations Targeting Deep Neural Networks on the Example of BrainScaleS-2
- 12.10.2023 - Vimala Bauer: A Deeply Threaded RISC-V Many-Core for FPGAs
- 25.09.2023 - Falk Selker: Optimisation of an Approximation Based Approach to Bayesian Neural Networks with TVM on Embedded Hardware
- 15.09.2023 - Klaus Fritz Breuer: Flex PCB Design for a Soft ExoGlove for Grasping Assistance
- 22.08.2023 - Patrick Szegedi: Design einer 60 V on-chip Ladungspumpe und eines Spannungsreglers zur Versorgung von SiPM Sensoren
- 04.08.2023 - Christian Alles: On the Performance of Butterfly Approximations on the Graphcore IPU
- 22.06.2023 - Daniel Barley: Reducing the state of large-scale MLPS by compressing the backward pass