Stefan Glüge

Annotation and Classification of Changes of Involvement in Group Conversation

The detection of involvement in a conversation is important to assess the level humans are participating in either a human-human or human-computer interaction. Especially, detecting changes in a groups involvement in a multi-party interaction is of interest to distinguish several constellations in the group itself. This information can further be used in situations where technical support of meetings is favoured, for instance, focusing a camera, switching microphones, etc. Moreover, this information could also help to improve the performance of technical systems applied in human-machine interaction. In this paper, we concentrate on video material given by the Table Talk corpus. Therefore, we introduce a way of annotating and classifying changes of involvement and discuss the reliability of the annotation. Further, we present classification results based on video features using Multi-Layer Networks.

Describing Human Emotions Through Mathematical Modelling

Abstract To design a companion technology we focus on the appraisal theory model to predict emotions and determine the appropriate system behaviour to support Human-Computer-Interaction. Until now, the implementation of emotion processing was hindered by the fact that the theories needed originate from diverging research areas, hence divergent research techniques and result representations are present. Since this difficulty arises repeatedly in interdisciplinary research, we investigated the use of mathematical modelling as an unifying language to translate the coherence of appraisal theory. We found that the mathematical category theory supports the modelling of human emotions according to the appraisal theory model and hence assists the implementation.

A Simple Recurrent Network for Implicit Learning of Temporal Sequences

A behavioural paradigm for learning arbitrary visuo-motor associations established that human observers learn to associate visual objects with their corresponding motor responses faster if the objects follow a temporal rule rather than if they were presented in a random order. Here, we use a simple recurrent network with a back propagation training algorithm adapted to a reinforcement learning scheme. Our simulations fit quantitatively as well as qualitatively to the behavioural results, endorsing the role of temporal context in associative learning scenarios.

Learning long-term dependencies in segmented-memory recurrent neural networks with backpropagation of error

Abstract In general, recurrent neural networks have difficulties in learning long-term dependencies. The segmented-memory recurrent neural network (SMRNN) architecture together with the extended real-time recurrent learning (eRTRL) algorithm was proposed to circumvent this problem. Due to its computational complexity eRTRL becomes impractical with increasing network size. Therefore, we introduce the less complex extended backpropagation through time (eBPTT) for SMRNN together with a layer-local unsupervised pre-training procedure. A comparison on the information latching problem showed that eRTRL is better able to handle the latching of information over longer periods of time, even though eBPTT guaranteed a better generalisation when training was successful. Further, pre-training significantly improved the ability to learn long-term dependencies with eBPTT. Therefore, the proposed eBPTT algorithm is suited for tasks that require big networks where eRTRL is impractical. The pre-training procedure itself is independent of the supervised learning algorithm and can improve learning in SMRNN in general.

Solving Number Series with Simple Recurrent Networks

Number series tests are a popular task in intelligence tests to measure a person’s ability of numerical reasoning. The function represented by a number series can be learned by artificial neural networks. In contrast to earlier research based on feedforward networks, we apply simple recurrent networks to the task of number series prediction. We systematically vary the number of input and hidden units in the networks to determine the optimal network configuration for the task. While feedforward networks could solve only 18 of 20 test series, a very small simple recurrent network could find a solution for all series. This underlines the importance of recurrence in such systems, which further is a basic concept in human cognition.

Dempster-Shafer theory with smoothness

This paper introduces the idea of a modified Dempster-Shafer theory. We adapt the belief characteristic of expert combination by introducing a penalty term which is specific to the investigated object. This approach is motivated by the observation that final decisions in the Dempster-Shafer theory might tend to fluctuations due to variations in sensor inputs on small time scales, even if the real phenomenological characteristic is stable.

Emotion recognition from speech using representation learning in extreme learning machines

We propose the use of an Extreme Learning Machine initialised as auto-encoder for emotion recognition from speech. This method is evaluated on three different speech corpora, namely EMO-DB, eNTERFACE and SmartKom. We compare our approach against state-of-the-art recognition rates achieved by Support Vector Machines (SVMs) and a deep learning approach based on Generalised Discriminant Analysis (GerDA). We could improve the recognition rate compared to SVMs by 3%–14% on all three corpora and those compared to GerDA by 8%–13% on two of the three corpora.

A markov model of conditional associative learning in a cognitive behavioural scenario

In conditional learning, one investigates the computational principles by which the human brain solves challenging recognition problems. The role of temporal context in the learning of arbitrary visuomotor associations has so far been studied mostly in primates. We model the explicit learning task where a sequence of visual objects is presented to human subjects. The computational modelling of the algorithms that appear to underlie human performance shall capture the effects of confusion in ordered and random presentation of objects. We present a Markov model where the learning history of a subject on a certain object is represented by the states of the model. The analysis of the resulting Markov chain makes it possible to judge the influence of two model parameters without the simulation of a specific learning scenario. As the model is able to reproduce the learning behaviour of human subjects it might be useful in the development of future experiments.