I offer topics in the field of Neural architecture search, mostly in performance prediction (but reach out if you have a specific topic in NAS or AutoML in mind). Depending on the results, the thesis could lead to a scientific publication.
Existing performance prediction works in NAS dealt with cell-based search spaces only - i.e., predicting the validation accuracy of a network based on its graph structure. However, in practice, we often vary other hyperparameters of the network, such as the number of filters, kernel size, etc., as well as parameters of the training (learning rate, batch size, SGD/Adam, etc.). Joint search spaces optimize both the architecture and hyperparameters of a network. The goal of this thesis is to propose novel methods for joint performance prediction and compare them with existing predictors.
Graph neural networks were successfully applied in performance prediction in NAS. However, a larger study on what graph neural architectures are the most suitable for performance prediction is missing. Similarly, different predictor training schemes could be used (e.g. binary relation predictor vs regressor). The goal of this thesis is to compare existing graph neural architectures and training schemes for performance prediction. It would be great to explore more recent GNNs that were not used in NAS performance prediction yet.