(HiWi) A student assistant for Investigating the impact of different sources of randomness in deep learning and developing methods to minimize their effects to enhance reproducibility:
The problem of reproducibility of deep learning models is a major concern and one of the reasons is uncontrolled randomization. To overcome this problem, better neural network training techniques that can reduce run-to-run variability are needed. Algorithmic model training for neural networks can pave the way, so we need to analyse it thoroughly. In algorithmic model training, the discrete weights are already distributed throughout the network and are selected based on quality scores. Unlike the conventional approach, algorithmic training never updates the weights but the quality scores.
- Understand the operation and source code of algorithmic model training.
- Conduct algorithmic model training experiments with different settings to analyse model variability with respect to the following factors:
- Different seed values
- Use of different hyperparameters (optimizer, learning rate)
- Number of weight choices per connection
- Number of trained models required to obtain useful results
- Time of convergence
- Extension of the algorithmic model to different architectures and datasets
- Bachelors/master's student studying computer science, applied mathematics, or a related subject.
- Expertise in Machine Learning, ideally in Deep Neural Networks (DNN): Convolutional Neural Networks (CNN) for image segmentation and classification.
- Proficiency in Python.
- Hands-on experience with deep learning toolkits like TensorFlow, PyTorch, etc.
- Ability to think analytically as well as strong communication skills
What we offer:
- Work on a real research project with practical orientation.
- Flexible working hours.
- 40/60 hours of work per month
- Possibility to work remotely, in-person, or hybrid.
Apply: Please send your complete application documents (curriculum vitae and current university transcript) via the email listed below by March 17th, 2023. Email: email@example.com OR firstname.lastname@example.org