(info) Information about writing a mastersthesis can be found here.

Archive

  • Sequential Recommender Systems
  • Modification and Testing of a SLAM Framework for Dynamic Environments
  • Multi-Agent Deep Reinforcement Learning
  • Correlation between physiological data and video features to better understand video induced emotions
  • Improvement of Deep-learned Object Detectors
  • Self Supervised Learning for Emotion Prediction induced by Videos
  • Machine Learning for Time Series Prediction in the Financial Industry
  • On Generating Mathematical Formulae
  • Detection of Teeth Grinding and Clenching using Surface Electromyography
  • Learned Under Water Feature Matcher based on Generated Artificial Image Base
  • Latency Prediction for Wireless System Synchronisation
  • Deep Learning for Financial Time Series Prediction
  • On the Assessment of RL
  • Federated Learning in Pedestrian Trajectory Prediction Tasks
  • Distilling Neural Networks for Real-Time Drone Control
  • Causal Regularization in Deep Learning Using the Average Causal Effect
    Student:Kathrin Khadra

    Abstract: Causal Interpretability aims to make decisions of algorithms interpretable by investigating what would have happened under different circumstances. These varying circumstances can be manipulations on the algorithm to assess its causality. In this thesis, I include the causal interpretability mechanism called the Average Causal Effect (ACE) into the training of a neural net. To assess the causality of the model, the ACE uses so-called interventions to manipulate the neural net. The goal is to determine more causal weights and biases during the model training. Using this approach, I evaluate whether including a causal interpretability mechanism as a regularization increases the overall causality of the model. Moreover, I investigate whether this improvement in causality also impacts the generalization ability of the neural net. The developed approach is compared to a standard neural net as well as L1 and L2 regularized neural nets. Furthermore, I conduct these experiments with well-balanced datasets, datasets with a prior probability shift, and datasets with a covariate shift. For all datasets, the results show that the presented causal regularization approach is able to improve the overall causality of the neural net. However, the distribution of the shifted training data highly affects the generalization ability. With an increasing variance of the distribution, the developed approach shows significantly lower test Mean Squared Errors than for training data with less variance. This is because the interventions applied by the ACE depend on the variance of the training data distribution.

    Email:-
    Status:

    FINISHED

    Supervisor:Matthias Kissel
  • Approximative Sparse Factorization of Neural Network Weight Matrices
    Student:Michael Brandner

    Abstract: In modern image processing applications, Convolutional Neural Networks (CNNs) are indispensable. Especially in the domains of object classification and face recognition, CNNs achieve impressive results. However, the increasingly accurate predictions are accompanied by ever-larger networks and consequently more computations. The number of operations required to compute the matrix vector product of a dense matrix M ∈ N N ×N in the fully connected layer scales with O (N 2 ) , mainly responsible that networks like VGG19 require 19.6 billion Floating-point Operations (FLOPs) to evaluate a single image. This thesis investigates the factorization of weight matrices, of the fully connected layer, into a product of sparse matrices, which potentially reduces the order of operations needed to the subquadratic domain. Consequently, the number of operations required for inference and thus, resource consumption is reduced. I examine three approximation algorithms, namely Butterfly factorization, sparse EigenGame, and Flexible Approximate MUlti-layer Sparse Transform ( F AμST ). The approaches are compared regarding the sparseness of their approximation and the approximation error. Furthermore, weight matrices of pre-trained Convolutional Neural Networks are factorized and compared regarding their prediction accuracy after approximation. 
    The best performance in terms of approximation error of the matrix and subsequent prediction accuracy was achieved by F AμST . F AμST was able to make sufficiently accurate predictions with only 20 % of the parameters. Where sufficient accurate means that the prediction accuracy drops by only 1 %. For similar results, the other algorithms needed 3 % (Butterfly) and 18 % (sparse EigenGame) more computations than the original matrix-vector product. 
    The experiments show that Approximative Sparse Factorization (ASF) of the weight matrices can significantly decrease resources consumption without deteriorating the accuracy of the predictions too much. This can enable complex computer vision algorithms to be used on devices with low computational resources or time-critical systems.

    Email:-
    Status:

    FINISHED

    Supervisor:Matthias Kissel
  • Algorithms for Matrix Approximations with Time Varying Systems
    Student:Stephan Nüßlein

    Abstract: There are different approaches to approximate matrices using structured matrices to reduce the computational cost of matrix-vector multiplications. A possible structure are sequentially semiseparable matrices, that describe the input-output behavior of time varying systems. If time varying systems are used to approximate weight matrices from neural networks, structural parameters have to be determined. In this thesis two algorithms to obtain the structural parameters are described. The first refines an initial segmentation by optimizing the input and output dimensions of the system. The second algorithm recursively splits the subsystems in a way that makes it possible to recover permuted sequentially semiseparable matrices. In experiments, the algorithms were able to recover the structure of simple test matrices. When used to approximate weight matrices from neural networks, the algorithms were able to reduce the computational cost of the matrix approximation compared to a naive approximation.

    Email:-
    Status:

    FINISHED

    Supervisor:Matthias Kissel

  • No labels