Seminare
Seminare des Lehrstuhls EIHW
Overview of Adaptive Learning-Rate-Based Optimisation Algorithms for Neural Networks
Description
Efficient learning algorithms are crucial for any machine learning algorithm in health care and other applications. A very crucial parameter for gradient-based optimisation algorithms is the learning rate, which determines how big changes in the network's parameters are within each optimisation step. Within the last few years, the concept of an adaptive learning rate has achieved great improvements in performance for learning algorithms.
Task
The task for this topic is to give an overview of the concepts, advantages, disadvantages and performances on different tasks of algorithms based on adapttive adaptive learning-rate-based RMSProp, ADADelta and Adam. For this purpose it will be necessary to search, read, understand and sum up the existing literature on the topic.
Utilises
Latex, Google Scholar
Requirements
Basic knowledge of neural networks, proficiency with calculus
Languages
German or English
Supervisor
Manuel Milling (manuel.milling@informatik.uni-augsburg.de)
Scent Challenge
Description
Scent plays an important role for human perception and can have a big impact on human behaviours. The ‘Scent Challenge’ provides a data corpus of human speech while exposed to five different smells.
Task
Classification of the speech data into the five classes of scents using Convolutional and Recurrent Neural Networks.
Utilises
Convolutional and Recurrent Neural Networks
Requirements
Preliminary machine learning knowledge needed
Languages
German or English
Supervisor
Shahin Amiriparian (shahin.amiriparian@informatik.uni-augsburg.de),
Manuel Milling (manuel.milling@informatik.uni-augsburg.de),
Thomas Wiest(thomas.wiest@informatik.uni-augsburg.de)
Loss Functions for Machine Learning Algorithms, literature review
Description
Efficient learning algorithms are crucial for any machine learning algorithm in health care and other applications. An important detail of any gradient-based optimisation algorithms is the loss function, which gives the algorithm its objective for optimisation.
Task
Give an overview of the concepts, advantages, disadvantages and performances on different tasks for various loss functions, such as mean squared error, cross entropy, and more complex loss functions including regularisation. For this purpose it will be necessary to search, read, understand and sum up the existing literature on the topic.
Utilises
Google Scholar
Requirements
Preliminary machine learning knowledge and calculus
Languages
German or English
Supervisor
Shahin Amiriparian (shahin.amiriparian@informatik.uni-augsburg.de),
Manuel Milling (manuel.milling@informatik.uni-augsburg.de),
Thomas Wiest(thomas.wiest@informatik.uni-augsburg.de)
Multilingual Sentiment Analysis in Text and Video Data, literature review
Description
Sentiment analysis is the process of computationally (using deep learning) identifying and categorizing opinions expressed by a writer or speaker, especially in order to determine whether the writer's attitude towards a particular topic, product, etc. is positive or negative. In recent years, a large number of multilingual representation learning in text has emerged, whereby a common representation of the used words from several languages is learned and creates an aligned vector space of word embeddings. This is particularly interesting if you have opinions (e.g. reviews) on a topic (e.g. a product) in different languages and can learn more variance across several languages through the aligned vectors.
Task
In this work, the student(s) will search and structure the latest research in the field of multilingual sentiment analysis that has utilised deep learning, either for text or video data. The aim is to identify existing literature and methods in the field of multilingual sentiment analysis in text and video.
Utilises
None, literature review only
Requirements
Preliminary knowledge in neural networks and representation learning (word embeddings)
Languages
German or English
Supervisor
Lukas Stappen, M. Sc. (lukas.stappen@informatik.uni-augsburg.de)
Advances in Few-Shot Learning on Text Data, literature review
Description
Deep learning approaches suffer from poor sampling efficiency in contrast to human perception - even a child could recognize an exotic animal when it sees a single image. One- and few-shot learning tries to learn representations from only a few samples and is often used in image recognition, such as facial recognition, where only few data and targets are available. Recently researchers started to use these techniques also in linguistic data. This is of particular interest, as a chronic lack of data in the field of deep learning in medicine is a constantly challenge.
Task
In this work, the student(s) will identify and structure the latest research in the field of n-shot learning, in particular, with a focus on topic, aspect detection, sentiment analysis and linguistic health data.
Utilises
None, literature review only
Requirements
Preliminary knowledge in neural networks and representation learning (word embeddings)
Languages
German or English
Supervisor
Lukas Stappen, M. Sc. (lukas.stappen@informatik.uni-augsburg.de)
Unsupervised Neural Topic Detection in Spoken Narratives, literature review
Description
Extracting the relevant topics and entities of a conversation is an important part of sentiment analysis that wants to categorize the opinions expressed towards a particular topic. Especially when designing data sets with the aim to make sentiment analysis supervised learnable, a prior extraction of the relevant topics is elementary, since only relevant entities and topics that are frequently used are important to annotate. A new type of neural network clustering emerged recently (https://www.aclweb.org/anthology/P17-1036) that showed good qualitative results and arouses interest in an evaluation of similar, more recently published work
Task
In this work, the student(s) identifies and structures the latest research in the field of unsupervised neural learning, with a particular focus on topic and aspect detection for sentiment analysis.
Utilises
None, literature review only
Requirements
Preliminary knowledge in neural networks and representation learning (word embeddings)
Languages
German or English
Supervisor
Lukas Stappen, M. Sc. (lukas.stappen@informatik.uni-augsburg.de)
Detecting Depression Patterns from Speech
Description
To diagnose depression, clinicians generally interview patients and identify the condition based on the patient's responses to specific questions. In recent years, machine learning models have been developed as a useful aid for diagnostics. For instance, the voices of depressed individuals reflects the perception of qualities such as monotony, slur, and less fluctuation. Hence, various previous works have focused on the correlation between depression and speech parameters.
Task
(a) survey previous works on the correlation between depression and speech parameters, and (b) hands-on evaluation on available depression databases.
Utilises
Python, scikit-learn, Tensorflow or PyTorch
Requirements
Programming skills in Python
Languages
English
Supervisor
Jing Han (jing.han@informatik.uni-augsburg.de)
Can emotions infer depression severity?
Description
The literature shows that paralinguistic information extracted from speech signals can be used to recognise depression severity in individuals. In this work, we go one step further, and we aim at exploring whether emotions are suitable indicators for this problem.
Task
Implement a system to recognise depression severity from speech signals, to be used as a baseline. Upgrade this system so it employs the emotions inferred from the speech signals as a predictor for depression severity. Finally, compare the performance of both systems.
Utilises
Python, PyTorch or Keras, scikit-Learn, openSMILE
Requirements
Good programming skills, previous knowledge in machine learning is a plus
Languages
English
Supervisor
Adria Mallol-Ragolta (adria.mallol-ragolta@informatik.uni-augsburg.de)
Explainable AI, literature review
Description
Machine learning have been applied in a lot of applications, such as emotion recognition, diagnosis of psychological/physiological diseases based on image/speech. However, the machine learning is like a black box for humans. To improve the recognition performance and give the patients more accurate diagnosis, we need to make the machine learning model explain its decision procedure.
Task
Reading explainable AI related articles/papers is the main task in this seminar. A final presentation is also required.
Utilises
-
Requirements
Knowledge of machine learning/deep learning theory
Languages
English
Supervisor
Zhao Ren (zhao.ren@informatik.uni-augsburg.de)
Sentiment analysis with hierarchical networks
Description
Hierarchical networks have been successfully employed in problems such as document classification or depression detection. In this work, we aim at using this type of networks to perform sentiment analysis from textual data. Due to the limitation of the available data for this problem in different languages, we also aim at analysing the performance of this system with automatically translated data.
Task
Implement a hierarchical network for sentiment analysis, and compare the performances obtained when using original and automatically translated data.
Utilises
Python, PyTorch and NLP techniques
Requirements
Good programming skills, previous knowledge in machine learning and NLP is a plus
Languages
English
Supervisor
Adria Mallol-Ragolta (adria.mallol-ragolta@informatik.uni-augsburg.de)
Chatting with my computer
Description
The popularity of chatbots is increasing, as they are able to simulate conversations. In this work, we aim at implementing a dialogue manager system able to have a conversation in a controlled scenario.
Task
Implementation of a chatbot in a scenario to be chosen by the student. The interaction can be computer-based using the keyboard as the interface.
Utilises
Python, PyTorch, scikit-learn
Requirements
Good programming skills, previous knowledge in machine learning is a plus
Languages
English
Supervisor
Adria Mallol-Ragolta (adria.mallol-ragolta@informatik.uni-augsburg.de),
Thomas Wiest (thomas.wiest@informatik.uni-augsburg.de)
Empathetic Behaviour Analysis with Deep Learning
Description
Empathic behaviour analysis is one of the most overlooked mechanisms in intelligent systems today. It can be defined as a complex process whereby "an observer reacting emotionally because he perceives that another is experiencing or about to experience an emotion”. Future machines should be endowed the ability to behave in an empathic manner, aiming at establishing and maintaining positive and long-term relationships with users.
Task
(a) study the work in the literature where empathetic behaviour is detected automatically in text/speech/video, (b) a preliminary hands-on evaluation of a deep learning approach for empathetic behaviour detection from speech and facial expressions.
Utilises
Python, Tensorflow or pytorch
Requirements
Programming skills in Python
Languages
English
Supervisor
Jing Han (jing.han@informatik.uni-augsburg.de)
Cross-Culture Emotion Recognition in the Wild
Description
Detecting and understanding emotional states of humans automatically is essential to improve the effectiveness of intelligent systems and devices, via providing an affective-based personalised user experience. Thanks to the great developments of machine learning, innovative technologies and algorithms are urging to handle affective information. Yet, in affective computing, there is not a lot of study on emotion recognition under cross-culture or multi-culture scenarios, by taking the effect of culture into account.
Task
(a) survey on the present techniques and related works in this emerging research field, and (b) evaluation on an audiovisual emotional dataset with 6 cultures.
Utilises
Python, Tensorflow or pytorch
Requirements
Preliminary knowledge of machine learning, good programming skills in Python
Languages
English
Supervisor
Jing Han (jing.han@informatik.uni-augsburg.de)
Unsupervised Neural Topic Detection in Spoken Narratives, literature review
Description
Extracting the relevant topics and entities of a conversation is an important part of sentiment analysis that wants to categorize the opinions expressed towards a particular topic. Especially when designing data sets with the aim to make sentiment analysis supervised learnable, a prior extraction of the relevant topics is elementary, since only relevant entities and topics that are frequently used are important to annotate. A new type of neural network clustering emerged recently (https://www.aclweb.org/anthology/P17-1036) that showed good qualitative results and arouses interest in an evaluation of similar, more recently published work
Task
In this work, the student(s) identifies and structures the latest research in the field of unsupervised neural learning, with a particular focus on topic and aspect detection for sentiment analysis.
Utilises
None, literature review only
Requirements
Preliminary knowledge of neural networks and representation learning (word embeddings)
Languages
German or English
Supervisor
Lukas Stappen, M. Sc. (lukas.stappen@informatik.uni-augsburg.de)
Advances in Few-Shot Learning on Text Data, literature review
Description
Deep learning approaches suffer from poor sampling efficiency in contrast to human perception - even a child could recognize an exotic animal when it sees a single image. One- and few-shot learning tries to learn representations from only a few samples and is often used in image recognition, such as facial recognition, where only few data and targets are available. Recently researchers started to use these techniques also in linguistic data. This is of particular interest, as a chronic lack of data in the field of deep learning in medicine is a constantly challenge.
Task
In this work, the student(s) will identify and structure the latest research in the field of n-shot learning, in particular, with a focus on topic, aspect detection, sentiment analysis and linguistic health data.
Utilises
None, literature review only
Requirements
Preliminary knowledge of neural networks and representation learning (word embeddings)
Languages
German or English
Supervisor
Lukas Stappen, M. Sc. (lukas.stappen@informatik.uni-augsburg.de)
Multilingual Sentiment Analysis in Text and Video Data, literature review
Description
Sentiment analysis is the process of computationally (using deep learning) identifying and categorizing opinions expressed by a writer or speaker, especially in order to determine whether the writer's attitude towards a particular topic, product, etc. is positive or negative. In recent years, a large number of multilingual representation learning in text has emerged, whereby a common representation of the used words from several languages is learned and creates an aligned vector space of word embeddings. This is particularly interesting if you have opinions (e.g. reviews) on a topic (e.g. a product) in different languages and can learn more variance across several languages through the aligned vectors.
Task
In this work, the student(s) will search and structure the latest research in the field of multilingual sentiment analysis that has utilised deep learning, either for text or video data. The aim is to identify existing literature and methods in the field of multilingual sentiment analysis in text and video.
Utilises
None, literature review only
Requirements
Preliminary knowledge of neural networks and representation learning (word embeddings)
Languages
German or English
Supervisor
Lukas Stappen, M. Sc. (lukas.stappen@informatik.uni-augsburg.de)
Overview of Adaptive Learning-Rate-Based Optimisation Algorithms for Neural Networks
Description
Efficient learning algorithms are crucial for any machine learning algorithm in health care and other applications. A very crucial parameter for gradient-based optimisation algorithms is the learning rate, which determines how big changes in the network's parameters are within each optimisation step. Within the last few years, the concept of an adaptive learning rate has achieved great improvements in performance for learning algorithms.
Task
The task for this topic is to give an overview of the concepts, advantages, disadvantages and performances on different tasks of algorithms based on adapttive adaptive learning-rate-based RMSProp, ADADelta and Adam. For this purpose it will be necessary to search, read, understand and sum up the existing literature on the topic.
Utilises
Latex, Google Scholar
Requirements
Basic knowledge of neural networks, proficiency with calculus
Languages
German or English
Supervisor
Manuel Milling (manuel.milling@informatik.uni-augsburg.de)
Seminare des Computer Audition
Augmentation of natural soundscapes
Description
In this seminar you will be exploring methods to extract musicality from natural sound environments, to augment the original data source.
Task
You will be provided with a dataset of emotional soundscapes (i.e. urban/natural/mechanical), and evaluate meaningful methods to extract musicality from the data (i.e. chroma features/ comb filters). You will then apply this to a synthesis engine and generate various styles of synthetic audio (i.e. sine tones/spherical tone/harmonics ). Finally performing an evaluation of the approaches chosen.
Utilises
Python - LibROSA/madmon
Requirements
Good programming experience in Python
Languages
English
Supervisor
Alice Baird (alice.baird@informatik.uni-augsburg.de)
Beat detection and generation from natural soundscapes
Description
In this seminar, you will be exploring methods to extract rhythm from natural sound environments, to augment the original data source.
Task
You will be provided with a dataset of emotional soundscapes (i.e. urban/ natural/ mechanical), and evaluate meaningful methods to extract temporal information from the data source. You will then apply this to a synthesis engine and generate various styles of drum beat (i.e electronic/ rock). Finally performing an evaluation of the approaches chosen.
Utilises
Python - LibROSA/madmon
Requirements
Good programming experience in Python
Languages
English
Supervisor
Alice Baird (alice.baird@informatik.uni-augsburg.de)
Analysis of Wellbeing during Audio Listening
Description
In this seminar, you will analyze a dataset gathered of individuals listening to various audio stimuli whilst having specific biological signals measured (including skin conductance, blood volume pressure).
Task
You will be provided with a dataset of individuals listening to various audio stimuli, and perform a series of statistical analysis on the data set to make conclusions in regards to the affect of audio on general wellbeing.
Utilises
Python - scikit-learn
Requirements
Programming experience in Python, basic knowledge of statistical analysis
Languages
English
Supervisor
Alice Baird (alice.baird@informatik.uni-augsburg.de)
Multimodal Deception Recognition 1: literature review
Description
Deception is an important tool in human-to-human interaction. Reliable recognition of deception gives people a big social advantage, but is nevertheless a very complex task. So far, only little effort has been made to develop automated deception recognition systems.
Task
Analysis of state-of-the-art for multimodal approaches for automated deception recognition.
Utilises
Google Scholar
Requirements
Preliminary machine learning knowledge helpful
Languages
German or English
Supervisor
Shahin Amiriparian (shahin.amiriparian@informatik.uni-augsburg.de),
Manuel Milling (manuel.milling@informatik.uni-augsburg.de),
Thomas Wiest(thomas.wiest@informatik.uni-augsburg.de)
Multimodal Deception Recognition 2: a novel dataset
Description
Deception is an important tool in human-to-human interaction. Reliable recognition of deception gives people a big social advantage, but is nevertheless a very complex task. So far, only little effort has been made to develop automated deception recognition systems.
Task
Preparation and conduction of an experimental setting for the collection of a novel deception dataset, which includes both audio and video modalities.
Utilises
Thermal Imaging Camera
Languages
German or English
Supervisor
Shahin Amiriparian (shahin.amiriparian@informatik.uni-augsburg.de),
Manuel Milling (manuel.milling@informatik.uni-augsburg.de),
Thomas Wiest(thomas.wiest@informatik.uni-augsburg.de)
Deep Speaker Identification
Description
A well working speaker identification system has a variety of applications, especially in the development of digital agents or mobile sensing applications. The problem of speaker identification however is a complicated task given a huge number of classes (one for each person), and in general only little data for every class. The VoxCeleb2 dataset provides 1 million utterances for 6,112 celebrities based on YouTube data.
Task
Training of a deep neural network architecture on the VoxCeleb2 dataset for speaker identification as a multi-class problem.
Utilises
Deep Neural Networks
Requirements
Preliminary machine learning knowledge
Languages
German or English
Supervisor
Shahin Amiriparian (shahin.amiriparian@informatik.uni-augsburg.de),
Manuel Milling (manuel.milling@informatik.uni-augsburg.de),
Thomas Wiest(thomas.wiest@informatik.uni-augsburg.de)
Pre-Trained Neural Networks for Audio Processing: A Survey, literature review
Description
Audio processing has a variety of applications, which all share abstract similarities given by the nature of audio data. Deep neural networks can be utilised to learn said similarities in form of features given big sets of data, such that individual applications only need to be fine-tuned to those abstract features.
Task
Analysis of state-of-the-art pre-trained neural networks for audio analysis.
Utilises
Google Scholar
Requirements
Preliminary machine learning knowledge helpful
Languages
German or English
Supervisor
Shahin Amiriparian (shahin.amiriparian@informatik.uni-augsburg.de),
Manuel Milling (manuel.milling@informatik.uni-augsburg.de),
Thomas Wiest (thomas.wiest@informatik.uni-augsburg.de)
Loss Functions for Machine Learning Algorithms. literature review
Description
Efficient learning algorithms are crucial for any machine learning algorithm in health care and other applications. An important detail of any gradient-based optimisation algorithms is the loss function, which gives the algorithm its objective for optimisation.
Task
Giving an overview of the concepts, advantages, disadvantages and performances on different tasks for various loss functions, such as mean squared error, cross entropy, and more complex loss functions including regularisation. For this purpose it will be necessary to search for, read, understand and sum up existing literature on the topic.
Utilises
Google Scholar
Requirements
Preliminary machine learning knowledge hand calculus
Languages
German or English
Supervisor
Shahin Amiriparian (shahin.amiriparian@informatik.uni-augsburg.de),
Manuel Milling (manuel.milling@informatik.uni-augsburg.de),
Thomas Wiest (thomas.wiest@informatik.uni-augsburg.de)
Speech Enhancement: Pursuit to Quality
Description & Task
Implement Neural Network architecture to enhance speech quality with big data (provided by EIHW). The experimental contents can be source separation (extract target speech from mixtures in multi-speakers case), or noise cancellation.
Utilises
Python, PyTorch (suggested) or TensorFlow1.0/2.0
Requirements
Background of Deep Learning, and Programming Experience with the above toolkits
Languages
English
Supervisor
Shuo Liu (shuo.liu@informatik.uni-augsburg.de)
Data Augmentation for Speech Emotion Recognition
Description
Deep networks are normally data-hungry, whereas collecting realistic emotional data is time-consuming and costly. In the absence of an adequate volume of training data, it is possible to increase the effective size of existing data via the process of data augmentation, which has contributed to significantly improving the performance of deep networks. In this case, it is essential to review and compare the effectiveness and reliability of various data augmentation approaches.
Task
(a) survey on existing conventional and advanced technologies for data augmentation, and/or (b) evaluate and compare 1-5 data augmentation approaches on an emotion database.
Utilises
Python, tensorflow or PyTorch
Requirements
Programming skills in Python
Languages
English
Supervisor
Jing Han (jing.han@informatik.uni-augsburg.de)
Sound Event Localization
Description
The detection of the environmental sound is a challenging task. In real life, different kinds of sound might occur at the same/different time, and keep with different time length. Different sounds might exist in the same time interval. Hence, sound event localization is important to better detect the sound.
Task
Reading related articles/papers and doing an experiment are the main tasks in this seminar.
Languages
English
Supervisor
Zhao Ren (zhao.ren@informatik.uni-augsburg.de)
A Survey of Pre-trained CNNs for Computer Audition
Description
How effective are pre-trained neural networks for computer audition tasks?
Task
An in-depth analysis of the state-of-the-art pre-trained CNNs
Utilises
-
Requirements
-
Languages
English
Supervisor
Shahin Amiriparian, M. Sc. (shahin.amiriparian@informatik.uni-augsburg.de)
Unsupervised Neural Topic Detection in Spoken Narratives, literature review
Description
Extracting the relevant topics and entities of a conversation is an important part of sentiment analysis that wants to categorize the opinions expressed towards a particular topic. Especially when designing data sets with the aim to make sentiment analysis supervised learnable, a prior extraction of the relevant topics is elementary, since only relevant entities and topics that are frequently used are important to annotate. A new type of neural network clustering emerged recently (https://www.aclweb.org/anthology/P17-1036) that showed good qualitative results and arouses interest in an evaluation of similar, more recently published work
Task
In this work, the student(s) identifies and structures the latest research in the field of unsupervised neural learning, with a particular focus on topic and aspect detection for sentiment analysis.
Utilises
None, literature review only
Requirements
Preliminary knowledge of neural networks and representation learning (word embeddings)
Languages
German or English
Supervisor
Lukas Stappen, M. Sc. (lukas.stappen@informatik.uni-augsburg.de)
Advances in Few-Shot Learning on Text Data, literature review
Description
Deep learning approaches suffer from poor sampling efficiency in contrast to human perception - even a child could recognize an exotic animal when it sees a single image. One- and few-shot learning tries to learn representations from only a few samples and is often used in image recognition, such as facial recognition, where only few data and targets are available. Recently researchers started to use these techniques also in linguistic data. This is of particular interest, as a chronic lack of data in the field of deep learning in medicine is a constantly challenge.
Task
In this work, the student(s) will identify and structure the latest research in the field of n-shot learning, in particular, with a focus on topic, aspect detection, sentiment analysis and linguistic health data.
Utilises
None, literature review only
Requirements
Preliminary knowledge of neural networks and representation learning (word embeddings)
Languages
German or English
Supervisor
Lukas Stappen, M. Sc. (lukas.stappen@informatik.uni-augsburg.de)
Multilingual Sentiment Analysis in Text and Video Data, literature review
Description
Sentiment analysis is the process of computationally (using deep learning) identifying and categorizing opinions expressed by a writer or speaker, especially in order to determine whether the writer's attitude towards a particular topic, product, etc. is positive or negative. In recent years, a large number of multilingual representation learning in text has emerged, whereby a common representation of the used words from several languages is learned and creates an aligned vector space of word embeddings. This is particularly interesting if you have opinions (e.g. reviews) on a topic (e.g. a product) in different languages and can learn more variance across several languages through the aligned vectors.
Task
In this work, the student(s) will search and structure the latest research in the field of multilingual sentiment analysis that has utilised deep learning, either for text or video data. The aim is to identify existing literature and methods in the field of multilingual sentiment analysis in text and video.
Utilises
None, literature review only
Requirements
Preliminary knowledge of neural networks and representation learning (word embeddings)
Languages
German or English
Supervisor
Lukas Stappen, M. Sc. (lukas.stappen@informatik.uni-augsburg.de)
Seminare der Sports Informatics
Effect of listening during sports activity, literature review
Description
In this seminar you will do a literature review on the topic of music and/or effect of sound stimuli during sports. Where possible, breaking this down into sports types, athletics, contact, team etc.
Task
You will perform a literature review on the topic of listening during sports, and also do an extensive search for datasets in this field which may have been gathered.
Languages
English
Supervisor
Alice Baird (alice.baird@informatik.uni-augsburg.de)
Noise-Cancelling in Sports-Related Acoustic Environments
Description
Noise-cancelling is a modern topic of research with various applications. We recently developed a new tool for noise-cancelling based on deep learning algorithms that has to be trained with different variations of noise and their according characteristics. So far we have not considered noise that results from sportive activity.
Task
Collection of a new audio data set containing noise originating from sportive activity and training of the noise-cancelling algorithm on the new data set.
Utilises
enHANS
Requirements
Basic programming knowledge helpful
Languages
German or English
Supervisor
Shuo Liu (shuo.liu@informatik.uni-augsburg.de)
Reminder of wrong decision in football games
Description
In a football game, the scores are usually given by a referee. Unfortunately, sometimes the referees might make mistakes, as the foul actions happen in a few seconds. This is too fast for human to give a correct decision. Hence, we propose to apply machine learning to construct a reminder for the referees when it is easy to have mistakes.
Task
Collecting video data from websites, and training a machine learning model.
Requirements
Image processing knowledge, machine learning knowledge, Python, and PyTorch
Languages
English
Supervisor
Zijiang Yang(zijiang.yang@informatik.uni-augsburg.de),
Zhao Ren (zhao.ren@informatik.uni-augsburg.de)
Speech-Based Heart Rate Monitoring
Description
Heart rate monitoring is an important task for various sport exercises, as well as medical conditions. This has led to a diverse range of technologies offering heart rate monitoring with varying accuracies. This seminar topic aims to evaluate the usability of speech to estimate heart rate.
Task
Preparation and conduction of an experimental setting for the collection of a speech and heart rate dataset in our lab in order to monitor heart rate based on speech.
Requirements
Sporty friends
Languages
English and German
Supervisor
Shahin Amiriparian (shahin.amiriparian@informatik.uni-augsburg.de),
Manuel Milling (manuel.milling@informatik.uni-augsburg.de),
Thomas Wiest (thomas.wiest@informatik.uni-augsburg.de)
Loss Functions for Machine Learning Algorithms, literature review
Description
Efficient learning algorithms are crucial for any machine learning algorithm in health care and other applications. An important detail of any gradient-based optimisation algorithms is the loss function, which gives the algorithm its objective for optimisation.
Task
Giving an overview of the concepts, advantages, disadvantages and performances on different tasks for various loss functions, such as mean squared error, cross entropy, and more complex loss functions including regularisation. For this purpose it will be necessary to search for, read, understand and sum up existing literature on the topic.
Utilises
Google Scholar
Requirements
Preliminary machine learning knowledge and calculus needed
Languages
English and German
Supervisor
Shahin Amiriparian (shahin.amiriparian@informatik.uni-augsburg.de),
Manuel Milling (manuel.milling@informatik.uni-augsburg.de),
Thomas Wiest (thomas.wiest@informatik.uni-augsburg.de)
Was that a score?
Description
The human voice carries a lot of information about the inner state, which can be important in several contexts. This topic aims to predict if a tennis player scores a point in a particular instance just based on their voice (groaning) and the sound of the racket hitting the ball.
Task
Collection of a new audio dataset of tennis players hitting the ball from YouTube, and baseline experiments using machine learning applications.
Utilises
auDeep, DeepSpectrum
Requirements
Basic programming knowledge helpful
Languages
English and German
Supervisor
Lukas Stappen (lukas.stappen@informatik.uni-augsburg.de),
Manuel Milling (manuel.milling@informatik.uni-augsburg.de)
Classification of Sports based on Audio
Description
Sound environments play an important role in everyday life. This topic is aiming to the classification of these sound environments, in particular for sports.
Task
Collection/extension of a new audio dataset including a wide range of sports from YouTube, and training a machine learning recognition system.
Utilises
CAS2T, auDeep, DeepSpectrum
Requirements
Basic programming knowledge helpful
Languages
English and German
Supervisor
Manuel Milling (manuel.milling@informatik.uni-augsburg.de)
Sports Ontology (Seminar/Thesis)
Description
To create one of the biggest sports datasets
Task
Collect a new dataset including a wide range of sports from YouTube using CAS²T and train a machine learning recognition system.
Utilises
CAS²T, auDeep, DeepSpectrum
Requirements
Basic programming knowledge
Languages
English and German
Supervisor
Shahin Amiriparian (shahin.amiriparian@informatik.uni-augsburg.de)