Design and development of new methodologies to detect and prevent overfitting in neural networks via noise propagation analysis

Umfeld

In the recent decade neural networks have become one of the key solutions for solving a range of machine vision problems. Neural network (NN) learns to capture the underlying relationship from the data using parameters or weights present in different layers of the NN. When a NN is either trained for a long period of time or has less data to learn from, it overfits to the training data distribution. This leads to a severe problem of not generalizing to new or unseen data, which degrades the performance of the network.

Aufgaben

The main aim of the thesis is to come up with solid theoretical and practical experiments which should be able to detect when a model overfits post training or during training. Related studies include the impact of training with noisier and adversarial data and analysing the inference performances on a neural network from Bayesian Deep Learning.

• Thorough analysis on the research papers focused specifically towards SOTA solutions to how and why a NN overfits.    
• Convert theoretical research into applied research towards the field of AV

Wir bieten

• An interdisciplinary working environment with partners from science, industry and users
• An economic/industrial work environment and organisation
• A pleasant working atmosphere and constructive cooperation

Wir erwarten

• Good programming skills in Python
• Good Theoretical knowledge in the field of machine learning or Deep learning
• Good understanding of Convolutional Neural Networks and Bayesian Neural Networks. RNN's and Reinforcement learning is a plus
• Ability to grasp research concepts at faster pace
• Any previous knowledge of ML frameworks such as Tensorflow, PyTorch is a plus
• Very good German or English language skills

Bewerbung

We're looking forward to receiving your PDF application to the mail vivekana@dont-want-spam.fzi.de or pavlitsk@dont-want-spam.fzi.de, with the following documents:

• Current excerpt of grade sheet
• Updated curriculum vitae