Finding State of the Art Techniques through Open Competitions

In our recent study group meeting, we discussed how a high learning rate is a regulator of neural network performance. Higher learning rates can result in better performing models. One aspect of the research discussed is that the first several epochs are strong indicators of the model performance.

Fast.ai has an open, accessible, image classification competition for finding better techniques using common GPUs. Students of Fast.ai have been inventing and testing new modeling techniques on an open competition. https://github.com/fastai/imagenette. The techniques identified in the competition have been applied to other visual models like segmentation and regression.

We'll discuss the competitions and the state of the art from the competitions.