SCCS Colloquium - Mar 4, 2020

From Sccswiki
Jump to navigation Jump to search
Date: March 4, 2020
Room: MI 01.09.014
Time: 15:00 - 16:00

Moritz Spielvogel: Active Learning on Classification Tasks

Master's thesis submission talk. Moritz is advised by Ionut Farcas, in collaboration with precibake.

Supervised machine learning based deep neural networks need in general big amounts of data to achieve a high accuracy on a test set. Depending on the task, it is very expensive to acquire such a big labeled data set whereas it is often easy to get huge amounts of unlabeled data. Active learning aims to construct from this pool of unlabeled data the most label-efficient data set by sampling iteratively subsets, which have to be labeled. Different active learning methods are evaluated in this thesis on two classification problems. Those active learning methods consist of uncertainty sampling, class balance sampling, representation sampling and regularization methods. For uncertainty sampling Bayesian neural network are a necessity, because they can model the network's uncertainty by placing a distribution over the network's parameters. On the data set provided by PreciBake, especially the uncertainty based sampling methods needed less than ten percent of the data set to achieve results equal to training on the whole data sets. Those methods have been improved by the combination with class balancing methods.

Keywords: Active Learning, Bayesian Neural Networks, Uncertainty Quanitfication, Class Balancy

Eric Fuchs: Comparison of distance metrics for MDS based NLDR using CNNs

Bachelor's thesis submission talk, in German. Eric is advised by Severin Reiz.

The L1 and L2 distance metrics can be used when training convolutional neural nets to perform nonlinear dimensionality reduction on image datasets, generating embedded spaces in a similar manner as with multidimensional scaling. The choice between them is often made arbitrarily. We trained enocder/decoder network pairs as Regressors, Autoencoders, Siamese networks, and with a triplet loss before applying them to Classification, Outlier detection, Interpolation, and Denoising. The experimental results were interpreted, subjectively where necessary, leading to the conclusion that using the euclidean or the manhattan distance during training matters less than the choice of training configuration. The L2 distance appeared minimally favorable.

Keywords: machine learning, dimensionality reduction