no code implementations • 29 Sep 2023 • Lukas Meiner, Jens Mehnert, Alexandru Paul Condurache
In particular, we are able to instantly drop 46. 72% of FLOPs while only losing 1. 25% accuracy by just swapping the convolution modules in a ResNet34 on CIFAR-10 for our HASTE module.
no code implementations • 29 Jun 2023 • Rinor Cakaj, Jens Mehnert, Bin Yang
However, we show experimentally that, despite the approximate additive penalty of BN, feature maps in deep neural networks (DNNs) tend to explode at the beginning of the network and that feature maps of DNNs contain large values during the whole training.
no code implementations • 29 Jun 2023 • Rinor Cakaj, Jens Mehnert, Bin Yang
Large weights in deep neural networks are a sign of a more complex network that is overfitted to the training data.
no code implementations • 17 May 2022 • Paul Wimmer, Jens Mehnert, Alexandru Paul Condurache
By freezing weights, the number of trainable parameters is shrunken which reduces gradient computations and the dimensionality of the model's optimization space.
no code implementations • CVPR 2022 • Paul Wimmer, Jens Mehnert, Alexandru Paul Condurache
Finally, we show that advances of IP are due to improved trainability and superior generalization ability.
no code implementations • 27 Jul 2021 • Paul Wimmer, Jens Mehnert, Alexandru Condurache
The combinatorial optimization problem given by COPS is relaxed on a linear program (LP).
no code implementations • 28 Nov 2020 • Paul Wimmer, Jens Mehnert, Alexandru Condurache
On the classification tasks MNIST and CIFAR-10/100 we outperform SNIP, in this setting the best reported one-shot pruning method, applied before training.