Sharad Kakran, TU Kaiserslautern
Generalization of Neural Predictors in Neural Architecture Search, Applied to Computer Vision
In recent years, the research topic “Neural Architecture Search” has gained a lot of popularity across the research community and industry, which mostly deals with searching for an architecture under predefined constraints for a task. Although in recent years, many sophisticated methods have been proposed to reduce the search time for exploring a search space, evaluation of so many architectures searched from a search space is computationally infeasible and this motivates the use of neural predictor which predicts the accuracy of searched architectures. However, the predictions of neural predictor strongly depend on how well it is trained which requires high quality of architecture-accuracy pairs, therefore involves a time and resource consuming process of training many sampled architectures from each new search space. Architectures, such as PSPNet, SSD, used in other computer vision tasks like semantic segmentation, object detection, share the same backbones from conventional image classification architectures, Resnet, VGG, therefore the search spaces across tasks overlap to some degree with common set of operations. Therefore we wish to investigate the question if knowledge from one task can be carried to another by using a neural predictor. A well trained neural predictor should learn a good embedding by encapsulating the information about neural architectures present in the search space such as how one operation affects another, what sets of operations result in good accuracy. We wish to evaluate how well the predictor generalises to other computer vision tasks with trained architectures sampled from the same search space.
How to join
The talk is held online via Zoom. You can join with the following link: