PhD Studentat IMLA
Home PageEmailGoogle ScholarTwitter
PhD Student working on the optimization of the intialization process of convolutional neural networks for computer vision tasks. In particular, I would like to replace the currently practiced i.i.d. initialization of convolution weights (filters) with a more sophisticated initialization, that leverages universal knowledge about converged filters. Ideally, this will allow faster convergence of training with lower energy consumption, training with fewer training samples, and potentially even a universal training-free feature extraction for common vision tasks.
Paul Gavrikov's papers