icon-symbol-logout-darkest-grey

News zum MScTI

Master Kolloquien

Lisa Kuhn (5.12.2023, 17:00, INF368 / Raum 524):
Scalability of Bayesian Neural Network Inference Methods for Real-World Tasks

Bayesian neural networks (BNNs), which combine the abilities of standard neural networks and Bayesian inference, provide a robust framework to reason reliably about uncertainty in predictions. Hence, they are a promising approach to overcome some of the key challenges of standard neural networks (NNs), namely overconfidence and sensitivity to overfitting. Since BNNs place distributions over weights, training them is more complex and resource-intensive compared to the already expensive training of standard NNs. Therefore, in order to train BNNs, several algorithms of varying complexity and based on different assumptions have emerged. In this thesis, we provide a detailed comparison between such algorithms concerning quality and scalability to growing problems and networks. With this, we aim to shed light on the relationship between quality and the resulting cost for BNNs. Regarding quality, our interest lies especially in the uncertainty estimation and ability to disentangle uncertainty into an aleatoric (inherent in the input) and an epistemic (from the model weights) component. To characterize the scalability of each algorithm, we make use of abstract hardware metrics such as the number of arithmetic operations and memory usage, as the runtime is highly specific to a given processor. On an exemplary regression and classification task, we show that Markov Chain Monte Carlo (MCMC) is able to provide superior results at the cost of higher memory consumption and significantly more arithmetic operations than alternative approaches. Mean-field Stochastic Variational Inference (SVI) and especially Monte Carlo Dropout (MC-Dropout) are efficient alternatives, but in our tests both provide less optimal uncertainty estimates and entangled uncertainties of varying degrees. Unfortunately, as dataset and network sizes grow, running MCMC quickly becomes infeasible, making SVI and MC-Dropout practical alternatives. This work contributes to the understanding of the trade-off between quality and computational costs of different BNN algorithms for practitioners. Consequently, our findings help to choose the most suitable BNN inference method for given quality constraints and hardware limitations.

Ibrahima Kouruma (12.12.2023, 10:00, INF 350 (OMZ) / Raum U012):
Design and development of a portable sensory system for biomechanical evaluation and motion-intention detection for lower-limb exoskeletons

From rigid lower-limb exoskeletons to soft exosuits for isolated joints, wearable devices are developed to assist individuals with walking disabilities in carrying out everyday activities. With a rising elderly population, demographic changes suggest that these type of devices will be more than a necessity in the near future. While technological advancements on the hardware and software of exoskeletons have seen a rapid increase over the last decade, the true impact and comparative support these devices offer, is not yet thoroughly investigated. To this end, this study proposes the design and development of a two-part portable sensory system that can be attached to a variety of exoskeletons, in order to evaluate the user’s performance and analyse gait. Through sensorised insoles and instrumented crutches, various biomechanical metrics can be measured and an algorithm for estimating gait phase and segmenting gait is proposed, which subsequently can serve for exoskeleton control in future work on this study.

Stipendien - Infoveranstaltung

Al 29.11.2023 organisiert die zentrale Studienberatung einen Infoabend zu Stipendien, s. Link.

Einführung WS23/24 und 'Thesis Fair'

Die Einführungsveranstaltung für unsere neuen Studierende des MScTI für das Wintersemester 2023/2024 fand am 16.10.2023 um 14:00 im Seminarraum des OMZ 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!

Geplante Veranstaltungen im WS23/24

Das Bild unten zeigt die (unverbindliche) Planung der Veranstaltungen im MScTI im kommenden WS23/24.

Wir möchten insbesondere auf die neuen Veranstaltungen hinweisen:

  • Architecture and CAD for FPGAs
  • Emerging Computing Paradigms
  • Digital Hardware Description and Verification
Geplante Veranstaltungen im MSc TI im WS23/24

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.

Einführungsveranstaltung MSc TI (SoSe 2023)

Die Einführungsveranstaltung für neue Studierende des MScTI für das Sommersemester 2023 fand am 17.04.2023 um 14:00 Uhr in INF 350 / OMZ R U014 statt.