Thesis supervisor: László Gulyás
Location of studies (in Hungarian): Eötvös Loránd University, Faculty of Informatics Abbreviation of location of studies: ELTE
Description of the research topic:
The original motivation of Artificial Neural Networks (ANNs) stems from biology: ANNs were modelled on our knowledge (at the time) of human neural networks.
Another important and hot topic of research of the past two decades is network science: the study of real world systems using the mathematical abstraction of a graph (network). An important lesson from network science is that networks from vastly different systems (e.g., biology, social systems, engineering, etc.) have important similarities, many of them share important structural properties and that useful statements can be made about their workings based on lessons learnt from networks from different domains, but with similar structural properties.
Among the many real world networks analyzed by network scientists, one can find works studying the neural networks of living organisms. These works suggest that real neural networks may have special structural properties, not unknown to network science, that may suggest that the network is a result of optimization.
This research connects network science to ANN, analyzing and optimizing the structure of ANNs, informed by network science. From the machine learning point of view, the project connects to the domain of optimization of ANN performance via sparsification and pruning, exploring the connections between the structure of sparse ANNs and their performance, taking into consideration both accuracy and cost (training speed).
Required language skills: English Recommended language skills (in Hungarian): Hungarian Further requirements: SW: python (Numpy, Pandas, Matplotlib), machine learning: TensorFlow, SKLearn, Pytorch, Keras, optional system development: Spark
Proactive, team-player student with good communication skills who is enthusiastic about data, cooperation ability