It can be challenging to train multi-task neural networks that outperform or even match their single-task counterparts. To help address this, we propose using knowledge distillation where single-task models teach a multi-task model. We enhance this training with teacher annealing, a novel method that gradually transitions the model from distillation to supervised learning, helping the multi-task model surpass its single-task teachers. We evaluate our approach by multi-task fine-tuning BERT on the GLUE benchmark. Our method consistently improves over standard single-task and multi-task training.
Recommendation problems with large numbers of discrete items, such as products, webpages, or videos, are ubiquitous in the technology industry. Deep neural networks are being increasingly used for these recommendation problems. These models use embeddings to represent discrete items as continuous vectors, and the vocabulary sizes and embedding dimensions, although heavily influence the model's accuracy, are often manually selected in a heuristical manner. We present Neural Input Search (NIS), a technique for learning the optimal vocabulary sizes and embedding dimensions for categorical features. The goal is to maximize prediction accuracy subject to a constraint on the total memory used by all embeddings. Moreover, we argue that the traditional Single-size Embedding (SE), which uses the same embedding dimension for all values of a feature, suffers from inefficient usage of model capacity and training data. We propose a novel type of embedding, namely Multi-size Embedding (ME), which allows the embedding dimension to vary for different values of the feature. During training we use reinforcement learning to find the optimal vocabulary size for each feature and embedding dimension for each value of the feature. In experiments on two common types of large scale recommendation problems, i.e. retrieval and ranking problems, NIS automatically found better vocabulary and embedding sizes that result in $6.8\%$ and $1.8\%$ relative improvements on Recall@1 and ROC-AUC over manually optimized ones.
Data augmentation is a critical component of training deep learning models. Although data augmentation has been shown to significantly improve image classification, its potential has not been thoroughly investigated for object detection. Given the additional cost for annotating images for object detection, data augmentation may be of even greater importance for this computer vision task. In this work, we study the impact of data augmentation on object detection. We first demonstrate that data augmentation operations borrowed from image classification may be helpful for training detection models, but the improvement is limited. Thus, we investigate how learned, specialized data augmentation policies improve generalization performance for detection models. Importantly, these augmentation policies only affect training and leave a trained model unchanged during evaluation. Experiments on the COCO dataset indicate that an optimized data augmentation policy improves detection accuracy by more than +2.3 mAP, and allow a single inference model to achieve a state-of-the-art accuracy of 50.7 mAP. Importantly, the best policy found on COCO may be transferred unchanged to other detection datasets and models to improve predictive accuracy. For example, the best augmentation policy identified with COCO improves a strong baseline on PASCAL-VOC by +2.7 mAP. Our results also reveal that a learned augmentation policy is superior to state-of-the-art architecture regularization methods for object detection, even when considering strong baselines. Code for training with the learned policy is available online at https://github.com/tensorflow/tpu/tree/master/models/official/detection
With the capability of modeling bidirectional contexts, denoising autoencoding based pretraining like BERT achieves better performance than pretraining approaches based on autoregressive language modeling. However, relying on corrupting the input with masks, BERT neglects dependency between the masked positions and suffers from a pretrain-finetune discrepancy. In light of these pros and cons, we propose XLNet, a generalized autoregressive pretraining method that (1) enables learning bidirectional contexts by maximizing the expected likelihood over all permutations of the factorization order and (2) overcomes the limitations of BERT thanks to its autoregressive formulation. Furthermore, XLNet integrates ideas from Transformer-XL, the state-of-the-art autoregressive model, into pretraining. Empirically, XLNet outperforms BERT on 20 tasks, often by a large margin, and achieves state-of-the-art results on 18 tasks including question answering, natural language inference, sentiment analysis, and document ranking.
Convolutional Neural Networks (ConvNets) are commonly developed at a fixed resource budget, and then scaled up for better accuracy if more resources are available. In this paper, we systematically study model scaling and identify that carefully balancing network depth, width, and resolution can lead to better performance. Based on this observation, we propose a new scaling method that uniformly scales all dimensions of depth/width/resolution using a simple yet highly effective compound coefficient. We demonstrate the effectiveness of this method on scaling up MobileNets and ResNet. To go even further, we use neural architecture search to design a new baseline network and scale it up to obtain a family of models, called EfficientNets, which achieve much better accuracy and efficiency than previous ConvNets. In particular, our EfficientNet-B7 achieves state-of-the-art 84.4% top-1 / 97.1% top-5 accuracy on ImageNet, while being 8.4x smaller and 6.1x faster on inference than the best existing ConvNet. Our EfficientNets also transfer well and achieve state-of-the-art accuracy on CIFAR-100 (91.7%), Flowers (98.8%), and 3 other transfer learning datasets, with an order of magnitude fewer parameters. Source code is at https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet.
We introduce a pretraining technique called Selfie, which stands for SELF-supervised Image Embedding. Selfie generalizes the concept of masked language modeling to continuous data, such as images. Given masked-out patches in an input image, our method learns to select the correct patch, among other "distractor" patches sampled from the same image, to fill in the masked location. This classification objective sidesteps the need for predicting exact pixel values of the target patches. The pretraining architecture includes a network of convolutional blocks to process patches followed by an attention pooling network to summarize the content of unmasked patches before predicting masked ones. During finetuning, we reuse the convolutional weights found by pretraining. We evaluate our method on three benchmarks (CIFAR-10, ImageNet 32 x 32, and ImageNet 224 x 224) with varying amounts of labeled data, from 5% to 100% of the training sets. Our pretraining method provides consistent improvements to ResNet-50 across all settings compared to the standard supervised training of the same network. Notably, on ImageNet 224 x 224 with 60 examples per class (5%), our method improves the mean accuracy of ResNet-50 from 35.6% to 46.7%, an improvement of 11.1 points in absolute accuracy. Our pretraining method also improves ResNet-50 training stability, especially on low data regime, by significantly lowering the standard deviation of test accuracies across datasets.
We present the next generation of MobileNets based on a combination of complementary search techniques as well as a novel architecture design. MobileNetV3 is tuned to mobile phone CPUs through a combination of hardware aware network architecture search (NAS) complemented by the NetAdapt algorithm and then subsequently improved through novel architecture advances. This paper starts the exploration of how automated search algorithms and network design can work together to harness complementary approaches improving the overall state of the art. Through this process we create two new MobileNet models for release: MobileNetV3-Large and MobileNetV3-Small which are targeted for high and low resource use cases. These models are then adapted and applied to the tasks of object detection and semantic segmentation. For the task of semantic segmentation (or any dense pixel prediction), we propose a new efficient segmentation decoder Lite Reduced Atrous Spatial Pyramid Pooling (LR-ASPP). We achieve new state of the art results for mobile classification, detection and segmentation. MobileNetV3-Large is 3.2% more accurate on ImageNet classification while reducing latency by 15% compared to MobileNetV2. MobileNetV2-Small is 4.6% more accurate while reducing latency by 5% compared to MobileNetV2. MobileNetV3-Large detection is 25% faster at roughly the same accuracy as MobileNetV2 on COCO detection. MobileNetV3-Large LR-ASPP is 30% faster than MobileNetV2 R-ASPP at similar accuracy for Cityscapes segmentation.
We investigate how the final parameters found by stochastic gradient descent are influenced by over-parameterization. We generate families of models by increasing the number of channels in a base network, and then perform a large hyper-parameter search to study how the test error depends on learning rate, batch size, and network width. We find that the optimal SGD hyper-parameters are determined by a "normalized noise scale," which is a function of the batch size, learning rate, and initialization conditions. In the absence of batch normalization, the optimal normalized noise scale is directly proportional to width. Wider networks, with their higher optimal noise scale, also achieve higher test accuracy. These observations hold for MLPs, ConvNets, and ResNets, and for two different parameterization schemes ("Standard" and "NTK"). We observe a similar trend with batch normalization for ResNets. Surprisingly, since the largest stable learning rate is bounded, the largest batch size consistent with the optimal normalized noise scale decreases as the width increases.
Despite its success, deep learning still needs large labeled datasets to succeed. Data augmentation has shown much promise in alleviating the need for more labeled data, but it so far has mostly been applied in supervised settings and achieved limited gains. In this work, we propose to apply data augmentation to unlabeled data in a semi-supervised learning setting. Our method, named Unsupervised Data Augmentation or UDA, encourages the model predictions to be consistent between an unlabeled example and an augmented unlabeled example. Unlike previous methods that use random noise such as Gaussian noise or dropout noise, UDA has a small twist in that it makes use of harder and more realistic noise generated by state-of-the-art data augmentation methods. This small twist leads to substantial improvements on six language tasks and three vision tasks even when the labeled set is extremely small. For example, on the IMDb text classification dataset, with only 20 labeled examples, UDA outperforms the state-of-the-art model trained on 25,000 labeled examples. On standard semi-supervised learning benchmarks, CIFAR-10 with 4,000 examples and SVHN with 1,000 examples, UDA outperforms all previous approaches and reduces more than $30\%$ of the error rates of state-of-the-art methods: going from 7.66% to 5.27% and from 3.53% to 2.46% respectively. UDA also works well on datasets that have a lot of labeled data. For example, on ImageNet, with 1.3M extra unlabeled data, UDA improves the top-1/top-5 accuracy from 78.28/94.36% to 79.04/94.45% when compared to AutoAugment.