Hanxiao Liu

David R. So, Wojciech Mańke, Hanxiao Liu et al.

Large Transformer models have been central to recent advances in natural language processing. The training and inference costs of these models, however, have grown rapidly and become prohibitively expensive. Here we aim to reduce the costs of Transformers by searching for a more efficient variant. Compared to previous approaches, our search is performed at a lower level, over the primitives that define a Transformer TensorFlow program. We identify an architecture, named Primer, that has a smaller training cost than the original Transformer and other variants for auto-regressive language modeling. Primer's improvements can be mostly attributed to two simple modifications: squaring ReLU activations and adding a depthwise convolution layer after each Q, K, and V projection in self-attention. Experiments show Primer's gains over Transformer increase as compute scale grows and follow a power law with respect to quality at optimal model sizes. We also verify empirically that Primer can be dropped into different codebases to significantly speed up training without additional tuning. For example, at a 500M parameter size, Primer improves the original T5 architecture on C4 auto-regressive language modeling, reducing the training cost by 4X. Furthermore, the reduced training cost means Primer needs much less compute to reach a target one-shot performance. For instance, in a 1.9B parameter configuration similar to GPT-3 XL, Primer uses 1/3 of the training compute to achieve the same one-shot performance as Transformer. We open source our models and several comparisons in T5 to help with reproducibility.

Gabriel Bender, Hanxiao Liu, Bo Chen et al.

Efficient Neural Architecture Search methods based on weight sharing have shown good promise in democratizing Neural Architecture Search for computer vision models. There is, however, an ongoing debate whether these efficient methods are significantly better than random search. Here we perform a thorough comparison between efficient and random search methods on a family of progressively larger and more challenging search spaces for image classification and detection on ImageNet and COCO. While the efficacies of both methods are problem-dependent, our experiments demonstrate that there are large, realistic tasks where efficient search methods can provide substantial gains over random search. In addition, we propose and evaluate techniques which improve the quality of searched architectures and reduce the need for manual hyper-parameter tuning. Source code and experiment data are available at https://github.com/google-research/google-research/tree/master/tunas

Bhuwan Dhingra, Hanxiao Liu, Zhilin Yang et al.

In this paper we study the problem of answering cloze-style questions over documents. Our model, the Gated-Attention (GA) Reader, integrates a multi-hop architecture with a novel attention mechanism, which is based on multiplicative interactions between the query embedding and the intermediate states of a recurrent neural network document reader. This enables the reader to build query-specific representations of tokens in the document for accurate answer selection. The GA Reader obtains state-of-the-art results on three benchmarks for this task--the CNN \& Daily Mail news stories and the Who Did What dataset. The effectiveness of multiplicative interaction is demonstrated by an ablation study, and by comparing to alternative compositional operators for implementing the gated-attention. The code is available at https://github.com/bdhingra/ga-reader.

Yanqi Zhou, Tao Lei, Hanxiao Liu et al.

Sparsely-activated Mixture-of-experts (MoE) models allow the number of parameters to greatly increase while keeping the amount of computation for a given token or a given sample unchanged. However, a poor expert routing strategy (e.g. one resulting in load imbalance) can cause certain experts to be under-trained, leading to an expert being under or over-specialized. Prior work allocates a fixed number of experts to each token using a top-k function regardless of the relative importance of different tokens. To address this, we propose a heterogeneous mixture-of-experts employing an expert choice method. Instead of letting tokens select the top-k experts, we have experts selecting the top-k tokens. As a result, each token can be routed to a variable number of experts and each expert can have a fixed bucket size. We systematically study pre-training speedups using the same computational resources of the Switch Transformer top-1 and GShard top-2 gating of prior work and find that our method improves training convergence time by more than 2x. For the same computational cost, our method demonstrates higher performance in fine-tuning 11 selected tasks in the GLUE and SuperGLUE benchmarks. For a smaller activation cost, our method outperforms the T5 dense model in 7 out of the 11 tasks.

Hieu Pham, Zihang Dai, Golnaz Ghiasi et al.

We present a combined scaling method - named BASIC - that achieves 85.7% top-1 accuracy on the ImageNet ILSVRC-2012 validation set without learning from any labeled ImageNet example. This accuracy surpasses best published similar models - CLIP and ALIGN - by 9.3%. Our BASIC model also shows significant improvements in robustness benchmarks. For instance, on 5 test sets with natural distribution shifts such as ImageNet-{A,R,V2,Sketch} and ObjectNet, our model achieves 84.3% top-1 average accuracy, only a small drop from its original ImageNet accuracy. To achieve these results, we scale up the contrastive learning framework of CLIP and ALIGN in three dimensions: data size, model size, and batch size. Our dataset has 6.6B noisy image-text pairs, which is 4x larger than ALIGN, and 16x larger than CLIP. Our largest model has 3B weights, which is 3.75x larger in parameters and 8x larger in FLOPs than ALIGN and CLIP. Finally, our batch size is 65536 which is 2x more than CLIP and 4x more than ALIGN. We encountered two main challenges with the scaling rules of BASIC. First, the main challenge with implementing the combined scaling rules of BASIC is the limited memory of accelerators, such as GPUs and TPUs. To overcome the memory limit, we propose two simple methods which make use of gradient checkpointing and model parallelism. Second, while increasing the dataset size and the model size has been the defacto method to improve the performance of deep learning models like BASIC, the effect of a large contrastive batch size on such contrastive-trained image-text models is not well-understood. To shed light on the benefits of large contrastive batch sizes, we develop a theoretical framework which shows that larger contrastive batch sizes lead to smaller generalization gaps for image-text models such as BASIC.

Zihang Dai, Hanxiao Liu, Quoc V. Le et al.

Transformers have attracted increasing interests in computer vision, but they still fall behind state-of-the-art convolutional networks. In this work, we show that while Transformers tend to have larger model capacity, their generalization can be worse than convolutional networks due to the lack of the right inductive bias. To effectively combine the strengths from both architectures, we present CoAtNets(pronounced "coat" nets), a family of hybrid models built from two key insights: (1) depthwise Convolution and self-Attention can be naturally unified via simple relative attention; (2) vertically stacking convolution layers and attention layers in a principled way is surprisingly effective in improving generalization, capacity and efficiency. Experiments show that our CoAtNets achieve state-of-the-art performance under different resource constraints across various datasets: Without extra data, CoAtNet achieves 86.0% ImageNet top-1 accuracy; When pre-trained with 13M images from ImageNet-21K, our CoAtNet achieves 88.56% top-1 accuracy, matching ViT-huge pre-trained with 300M images from JFT-300M while using 23x less data; Notably, when we further scale up CoAtNet with JFT-3B, it achieves 90.88% top-1 accuracy on ImageNet, establishing a new state-of-the-art result.

Hanxiao Liu, Zihang Dai, David R. So et al.

Transformers have become one of the most important architectural innovations in deep learning and have enabled many breakthroughs over the past few years. Here we propose a simple network architecture, gMLP, based on MLPs with gating, and show that it can perform as well as Transformers in key language and vision applications. Our comparisons show that self-attention is not critical for Vision Transformers, as gMLP can achieve the same accuracy. For BERT, our model achieves parity with Transformers on pretraining perplexity and is better on some downstream NLP tasks. On finetuning tasks where gMLP performs worse, making the gMLP model substantially larger can close the gap with Transformers. In general, our experiments show that gMLP can scale as well as Transformers over increased data and compute.

Daiyi Peng, Xuanyi Dong, Esteban Real et al.

Neural networks are sensitive to hyper-parameter and architecture choices. Automated Machine Learning (AutoML) is a promising paradigm for automating these choices. Current ML software libraries, however, are quite limited in handling the dynamic interactions among the components of AutoML. For example, efficientNAS algorithms, such as ENAS and DARTS, typically require an implementation coupling between the search space and search algorithm, the two key components in AutoML. Furthermore, implementing a complex search flow, such as searching architectures within a loop of searching hardware configurations, is difficult. To summarize, changing the search space, search algorithm, or search flow in current ML libraries usually requires a significant change in the program logic. In this paper, we introduce a new way of programming AutoML based on symbolic programming. Under this paradigm, ML programs are mutable, thus can be manipulated easily by another program. As a result, AutoML can be reformulated as an automated process of symbolic manipulation. With this formulation, we decouple the triangle of the search algorithm, the search space and the child program. This decoupling makes it easy to change the search space and search algorithm (without and with weight sharing), as well as to add search capabilities to existing code and implement complex search flows. We then introduce PyGlove, a new Python library that implements this paradigm. Through case studies on ImageNet and NAS-Bench-101, we show that with PyGlove users can easily convert a static program into a search space, quickly iterate on the search spaces and search algorithms, and craft complex search flows to achieve better results.

Yanqi Zhou, Sudip Roy, Amirali Abdolrashidi et al.

Most compilers for machine learning (ML) frameworks need to solve many correlated optimization problems to generate efficient machine code. Current ML compilers rely on heuristics based algorithms to solve these optimization problems one at a time. However, this approach is not only hard to maintain but often leads to sub-optimal solutions especially for newer model architectures. Existing learning based approaches in the literature are sample inefficient, tackle a single optimization problem, and do not generalize to unseen graphs making them infeasible to be deployed in practice. To address these limitations, we propose an end-to-end, transferable deep reinforcement learning method for computational graph optimization (GO), based on a scalable sequential attention mechanism over an inductive graph neural network. GO generates decisions on the entire graph rather than on each individual node autoregressively, drastically speeding up the search compared to prior methods. Moreover, we propose recurrent attention layers to jointly optimize dependent graph optimization tasks and demonstrate 33%-60% speedup on three graph optimization tasks compared to TensorFlow default optimization. On a diverse set of representative graphs consisting of up to 80,000 nodes, including Inception-v3, Transformer-XL, and WaveNet, GO achieves on average 21% improvement over human experts and 18% improvement over the prior state of the art with 15x faster convergence, on a device placement task evaluated in real systems.

Barret Zoph, Golnaz Ghiasi, Tsung-Yi Lin et al.

Pre-training is a dominant paradigm in computer vision. For example, supervised ImageNet pre-training is commonly used to initialize the backbones of object detection and segmentation models. He et al., however, show a surprising result that ImageNet pre-training has limited impact on COCO object detection. Here we investigate self-training as another method to utilize additional data on the same setup and contrast it against ImageNet pre-training. Our study reveals the generality and flexibility of self-training with three additional insights: 1) stronger data augmentation and more labeled data further diminish the value of pre-training, 2) unlike pre-training, self-training is always helpful when using stronger data augmentation, in both low-data and high-data regimes, and 3) in the case that pre-training is helpful, self-training improves upon pre-training. For example, on the COCO object detection dataset, pre-training benefits when we use one fifth of the labeled data, and hurts accuracy when we use all labeled data. Self-training, on the other hand, shows positive improvements from +1.3 to +3.4AP across all dataset sizes. In other words, self-training works well exactly on the same setup that pre-training does not work (using ImageNet to help COCO). On the PASCAL segmentation dataset, which is a much smaller dataset than COCO, though pre-training does help significantly, self-training improves upon the pre-trained model. On COCO object detection, we achieve 54.3AP, an improvement of +1.5AP over the strongest SpineNet model. On PASCAL segmentation, we achieve 90.5 mIOU, an improvement of +1.5% mIOU over the previous state-of-the-art result by DeepLabv3+.

Hanxiao Liu, Andrew Brock, Karen Simonyan et al.

Normalization layers and activation functions are fundamental components in deep networks and typically co-locate with each other. Here we propose to design them using an automated approach. Instead of designing them separately, we unify them into a single tensor-to-tensor computation graph, and evolve its structure starting from basic mathematical functions. Examples of such mathematical functions are addition, multiplication and statistical moments. The use of low-level mathematical functions, in contrast to the use of high-level modules in mainstream NAS, leads to a highly sparse and large search space which can be challenging for search methods. To address the challenge, we develop efficient rejection protocols to quickly filter out candidate layers that do not work well. We also use multi-objective evolution to optimize each layer's performance across many architectures to prevent overfitting. Our method leads to the discovery of EvoNorms, a set of new normalization-activation layers with novel, and sometimes surprising structures that go beyond existing design patterns. For example, some EvoNorms do not assume that normalization and activation functions must be applied sequentially, nor need to center the feature maps, nor require explicit activation functions. Our experiments show that EvoNorms work well on image classification models including ResNets, MobileNets and EfficientNets but also transfer well to Mask R-CNN with FPN/SpineNet for instance segmentation and to BigGAN for image synthesis, outperforming BatchNorm and GroupNorm based layers in many cases.

Jiahui Yu, Pengchong Jin, Hanxiao Liu et al.

Neural architecture search (NAS) has shown promising results discovering models that are both accurate and fast. For NAS, training a one-shot model has become a popular strategy to rank the relative quality of different architectures (child models) using a single set of shared weights. However, while one-shot model weights can effectively rank different network architectures, the absolute accuracies from these shared weights are typically far below those obtained from stand-alone training. To compensate, existing methods assume that the weights must be retrained, finetuned, or otherwise post-processed after the search is completed. These steps significantly increase the compute requirements and complexity of the architecture search and model deployment. In this work, we propose BigNAS, an approach that challenges the conventional wisdom that post-processing of the weights is necessary to get good prediction accuracies. Without extra retraining or post-processing steps, we are able to train a single set of shared weights on ImageNet and use these weights to obtain child models whose sizes range from 200 to 1000 MFLOPs. Our discovered model family, BigNASModels, achieve top-1 accuracies ranging from 76.5% to 80.9%, surpassing state-of-the-art models in this range including EfficientNets and Once-for-All networks without extra retraining or post-processing. We present ablative study and analysis to further understand the proposed BigNASModels.

Wei Wen, Hanxiao Liu, Hai Li et al.

Neural Architecture Search methods are effective but often use complex algorithms to come up with the best architecture. We propose an approach with three basic steps that is conceptually much simpler. First we train N random architectures to generate N (architecture, validation accuracy) pairs and use them to train a regression model that predicts accuracy based on the architecture. Next, we use this regression model to predict the validation accuracies of a large number of random architectures. Finally, we train the top-K predicted architectures and deploy the model with the best validation result. While this approach seems simple, it is more than 20 times as sample efficient as Regularized Evolution on the NASBench-101 benchmark and can compete on ImageNet with more complex approaches based on weight sharing, such as ProxylessNAS.

Yanqi Zhou, Sudip Roy, Amirali Abdolrashidi et al.

Runtime and scalability of large neural networks can be significantly affected by the placement of operations in their dataflow graphs on suitable devices. With increasingly complex neural network architectures and heterogeneous device characteristics, finding a reasonable placement is extremely challenging even for domain experts. Most existing automated device placement approaches are impractical due to the significant amount of compute required and their inability to generalize to new, previously held-out graphs. To address both limitations, we propose an efficient end-to-end method based on a scalable sequential attention mechanism over a graph neural network that is transferable to new graphs. On a diverse set of representative deep learning models, including Inception-v3, AmoebaNet, Transformer-XL, and WaveNet, our method on average achieves 16% improvement over human experts and 9.2% improvement over the prior art with 15 times faster convergence. To further reduce the computation cost, we pre-train the policy network on a set of dataflow graphs and use a superposition network to fine-tune it on each individual graph, achieving state-of-the-art performance on large hold-out graphs with over 50k nodes, such as an 8-layer GNMT.

Guokun Lai, Wei-Cheng Chang, Yiming Yang et al.

Multivariate time series forecasting is an important machine learning problem across many domains, including predictions of solar plant energy output, electricity consumption, and traffic jam situation. Temporal data arise in these real-world applications often involves a mixture of long-term and short-term patterns, for which traditional approaches such as Autoregressive models and Gaussian Process may fail. In this paper, we proposed a novel deep learning framework, namely Long- and Short-term Time-series network (LSTNet), to address this open challenge. LSTNet uses the Convolution Neural Network (CNN) and the Recurrent Neural Network (RNN) to extract short-term local dependency patterns among variables and to discover long-term patterns for time series trends. Furthermore, we leverage traditional autoregressive model to tackle the scale insensitive problem of the neural network model. In our evaluation on real-world data with complex mixtures of repetitive patterns, LSTNet achieved significant performance improvements over that of several state-of-the-art baseline methods. All the data and experiment codes are available online.

Hanxiao Liu, Yiming Yang

Cross-graph Relational Learning (CGRL) refers to the problem of predicting the strengths or labels of multi-relational tuples of heterogeneous object types, through the joint inference over multiple graphs which specify the internal connections among each type of objects. CGRL is an open challenge in machine learning due to the daunting number of all possible tuples to deal with when the numbers of nodes in multiple graphs are large, and because the labeled training instances are extremely sparse as typical. Existing methods such as tensor factorization or tensor-kernel machines do not work well because of the lack of convex formulation for the optimization of CGRL models, the poor scalability of the algorithms in handling combinatorial numbers of tuples, and/or the non-transductive nature of the learning methods which limits their ability to leverage unlabeled data in training. This paper proposes a novel framework which formulates CGRL as a convex optimization problem, enables transductive learning using both labeled and unlabeled tuples, and offers a scalable algorithm that guarantees the optimal solution and enjoys a linear time complexity with respect to the sizes of input graphs. In our experiments with a subset of DBLP publication records and an Enzyme multi-source dataset, the proposed method successfully scaled to the large cross-graph inference problem, and outperformed other representative approaches significantly.