resnext101_32x8d_wsl: the ConvNet pre-trained on Instagram hashtags and fine-tuned on ImageNet, yielding a record-breaking 85.4% top-1 accuracy is now available. https://t.co/5w04pugDcC
— Yann LeCun (@ylecun) June 23, 2019
resnext101_32x8d_wsl: the ConvNet pre-trained on Instagram hashtags and fine-tuned on ImageNet, yielding a record-breaking 85.4% top-1 accuracy is now available. https://t.co/5w04pugDcC
— Yann LeCun (@ylecun) June 23, 2019
I remember that DeepDream has a tendency to hallucinate dog pictures as it used VGG trained on ImageNet, which contained many dogs. Wonder what DeepDream with pre-trained model on Instagram would look like. https://t.co/mqOG5gBUYw pic.twitter.com/pB2UyN65si
— hardmaru (@hardmaru) June 24, 2019
1/3 This is a good study which shows the extent of diminishing returns in deep learning: https://t.co/LiU3ZFfkGE . The largest model has a whopping 829M parameters, gets 3% performance gain over a model with 10x less parameters, with flattening curve. pic.twitter.com/PlgQr1u8fr
— Filip Piekniewski (@filippie509) June 24, 2019
Karpathy suggested 2%/98% as a lower bound for ImageNet top-5: https://t.co/eT3tfnyTbZ EfficientNet is at 97.1% & the Instagram ResNet at 97.6%, so... (Not sure anyone's tried to estimate a lower bound for top-1.)
— 𝔊𝔴𝔢𝔯𝔫 (@gwern) June 25, 2019