Papers of the day   All papers

Compounding the Performance Improvements of Assembled Techniques in a Convolutional Neural Network

Comments

Jeremy Howard: This is a very interesting paper. It shows that a tweaked ResNet50 is about as accurate as EfficientNet-B4 but >3x faster. The EfficientNet paper measured FLOPS, which is a theoretical performance measure, rather than time, which is what actually matters. https://arxiv.org/abs/2001.06268 https://t.co/cyBiueqPuf

10 replies, 661 likes


Ross Wightman: @jeremyphoward All of the models in this family (MNASNet, FBNet, MobileNet-v3, EfficientNet) are pretty challenging and slow to train to spec'd accuracy. Epoch wise, finding the right hyper-params/techniques, and GPU memory use... tricks that work with ResNet, etc. also don't work as well here

1 replies, 12 likes


Daisuke Okanohara: A variety of techniques (architecture and regularization) have been proposed for improving CNNs. Assembling these techniques carefully to optimize accuracy and throughput enables ResNet to achieve similar accuracy as EfficientNet while 3x~10x throughput. https://arxiv.org/abs/2001.06268

0 replies, 11 likes


Yann Dauphin: Combining Mixup and other recent techniques brings good old Resnet-50 from 76.3% to 82.79% on Imagenet AND the network has state-of-the-art efficiency. https://arxiv.org/pdf/2001.06268.pdf https://t.co/MwbusraUZr

0 replies, 6 likes


那須音トウ | imenurok: Compounding the Performance Improvements of Assembled Techniques in a Convolutional Neural Network https://arxiv.org/abs/2001.06268

1 replies, 5 likes


Morgan McGuire: Resnet-50 up to 82.8% top-1 imagenet accuracy 🤯 combo of architecture tweaks and regularisation "Compounding the Performance Improvements of Assembled Techniques in a Convolutional Neural Network" - https://arxiv.org/pdf/2001.06268.pdf Hat-tip to fastai forums https://forums.fast.ai/t/fastai-v2-additional-models/62067/3

0 replies, 3 likes


akira: https://arxiv.org/abs/2001.06268 Build the network ,that combines some powerful existing techniques, achieve same accuracy as EfficientNet B6 + AutoAug but 5 times faster. The authors say that the latest ones such as AugMix are not used here, so the accuracy may still be improved. https://t.co/J83LVHVxbP

0 replies, 2 likes


arXiv CS-CV: Compounding the Performance Improvements of Assembled Techniques in a Convolutional Neural Network http://arxiv.org/abs/2001.06268

0 replies, 2 likes


arXiv CS-CV: Compounding the Performance Improvements of Assembled Techniques in a Convolutional Neural Network http://arxiv.org/abs/2001.06268

0 replies, 1 likes


Nicolas Won: Check out our new paper "Compounding the performance accuracy of assembled techniques in a convolutional neural network". Our proposed ResNet-50 shows an improvement in top-1 accuracy from 76.3% to 82.78%. https://arxiv.org/abs/2001.06268

1 replies, 0 likes


Content

Found on Jan 21 2020 at https://arxiv.org/pdf/2001.06268.pdf

PDF content of a computer science paper: Compounding the Performance Improvements of Assembled Techniques in a Convolutional Neural Network