Vcc: Scaling Transformers to 128K Tokens or More by Prioritizing Important Tokens

Transformer models are foundational to natural language processing (NLP) and computer vision. Despite various recent works devoted to reducing the quadratic cost of such models (as a function of the sequence length $n$), dealing with ultra long sequences efficiently (e.g., with more than 16K tokens) remains challenging. Applications such as answering questions based on an entire book or summarizing a scientific article are inefficient or infeasible. In this paper, we propose to significantly reduce the dependency of a Transformer model's complexity on $n$, by compressing the input into a representation whose size $r$ is independent of $n$ at each layer. Specifically, by exploiting the fact that in many tasks, only a small subset of special tokens (we call VIP-tokens) are most relevant to the final prediction, we propose a VIP-token centric compression (Vcc) scheme which selectively compresses the input sequence based on their impact on approximating the representation of these VIP-tokens. Compared with competitive baselines, the proposed algorithm not only is efficient (achieving more than $3\times$ efficiency improvement compared to baselines on 4K and 16K lengths), but also achieves competitive or better performance on a large number of tasks. Further, we show that our algorithm can be scaled to 128K tokens (or more) while consistently offering accuracy improvement.

Prompt Pre-Training with Twenty-Thousand Classes for Open-Vocabulary Visual Recognition

This work proposes POMP, a prompt pre-training method for vision-language models. Being memory and computation efficient, POMP enables the learned prompt to condense semantic information for a rich set of visual concepts with over twenty-thousand classes. Once pre-trained, the prompt with a strong transferable ability can be directly plugged into a variety of visual recognition tasks including image classification, semantic segmentation, and object detection, to boost recognition performances in a zero-shot manner. Empirical evaluation shows that POMP achieves state-of-the-art performances on 21 downstream datasets, e.g., 67.0% average accuracy on 10 classification dataset (+3.1% compared to CoOp) and 84.4 hIoU on open-vocabulary Pascal VOC segmentation (+6.9 compared to ZSSeg).

Disentangling Orthogonal Planes for Indoor Panoramic Room Layout Estimation with Cross-Scale Distortion Awareness

Based on the Manhattan World assumption, most existing indoor layout estimation schemes focus on recovering layouts from vertically compressed 1D sequences. However, the compression procedure confuses the semantics of different planes, yielding inferior performance with ambiguous interpretability. To address this issue, we propose to disentangle this 1D representation by pre-segmenting orthogonal (vertical and horizontal) planes from a complex scene, explicitly capturing the geometric cues for indoor layout estimation. Considering the symmetry between the floor boundary and ceiling boundary, we also design a soft-flipping fusion strategy to assist the pre-segmentation. Besides, we present a feature assembling mechanism to effectively integrate shallow and deep features with distortion distribution awareness. To compensate for the potential errors in pre-segmentation, we further leverage triple attention to reconstruct the disentangled sequences for better performance. Experiments on four popular benchmarks demonstrate our superiority over existing SoTA solutions, especially on the 3DIoU metric. The code is available at \url{https://github.com/zhijieshen-bjtu/DOPNet}.

Decoupled Model Schedule for Deep Learning Training

Recent years have seen an increase in the development of large deep learning (DL) models, which makes training efficiency crucial. Common practice is struggling with the trade-off between usability and performance. On one hand, DL frameworks such as PyTorch use dynamic graphs to facilitate model developers at a price of sub-optimal model training performance. On the other hand, practitioners propose various approaches to improving the training efficiency by sacrificing some of the flexibility, ranging from making the graph static for more thorough optimization (e.g., XLA) to customizing optimization towards large-scale distributed training (e.g., DeepSpeed and Megatron-LM). In this paper, we aim to address the tension between usability and training efficiency through separation of concerns. Inspired by DL compilers that decouple the platform-specific optimizations of a tensor-level operator from its arithmetic definition, this paper proposes a schedule language to decouple model execution from definition. Specifically, the schedule works on a PyTorch model and uses a set of schedule primitives to convert the model for common model training optimizations such as high-performance kernels, effective 3D parallelism, and efficient activation checkpointing. Compared to existing optimization solutions, we optimize the model as-needed through high-level primitives, and thus preserving programmability and debuggability for users to a large extent. Our evaluation results show that by scheduling the existing hand-crafted optimizations in a systematic way, we are able to improve training throughput by up to 3.35x on a single machine with 8 NVIDIA V100 GPUs, and by up to 1.32x on multiple machines with up to 64 GPUs, when compared to the out-of-the-box performance of DeepSpeed and Megatron-LM.

SPT: Semi-Parametric Prompt Tuning for Multitask Prompted Learning

Pre-trained large language models can efficiently interpolate human-written prompts in a natural way. Multitask prompted learning can help generalization through a diverse set of tasks at once, thus enhancing the potential for more effective downstream fine-tuning. To perform efficient multitask-inference in the same batch, parameter-efficient fine-tuning methods such as prompt tuning have been proposed. However, the existing prompt tuning methods may lack generalization. We propose SPT, a semi-parametric prompt tuning method for multitask prompted learning. The novel component of SPT is a memory bank from where memory prompts are retrieved based on discrete prompts. Extensive experiments, such as (i) fine-tuning a full language model with SPT on 31 different tasks from 8 different domains and evaluating zero-shot generalization on 9 heldout datasets under 5 NLP task categories and (ii) pretraining SPT on the GLUE datasets and evaluating fine-tuning on the SuperGLUE datasets, demonstrate effectiveness of SPT.

SMILE: Scaling Mixture-of-Experts with Efficient Bi-level Routing

The mixture of Expert (MoE) parallelism is a recent advancement that scales up the model size with constant computational cost. MoE selects different sets of parameters (i.e., experts) for each incoming token, resulting in a sparsely-activated model. Despite several successful applications of MoE, its training efficiency degrades significantly as the number of experts increases. The routing stage in MoE relies on the efficiency of the All2All communication collective, which suffers from network congestion and has poor scalability. To mitigate these issues, we introduce SMILE, which exploits heterogeneous network bandwidth and splits a single-step routing into bi-level routing. Our experimental results show that the proposed method obtains a 2.5x speedup over Switch Transformer in terms of pretraining throughput on the Colossal Clean Crawled Corpus without losing any convergence speed.

Node-oriented Spectral Filtering for Graph Neural Networks

Graph neural networks (GNNs) have shown remarkable performance on homophilic graph data while being far less impressive when handling non-homophilic graph data due to the inherent low-pass filtering property of GNNs. In general, since the real-world graphs are often a complex mixture of diverse subgraph patterns, learning a universal spectral filter on the graph from the global perspective as in most current works may still suffer from great difficulty in adapting to the variation of local patterns. On the basis of the theoretical analysis on local patterns, we rethink the existing spectral filtering methods and propose the \textbf{\underline{N}}ode-oriented spectral \textbf{\underline{F}}iltering for \textbf{\underline{G}}raph \textbf{\underline{N}}eural \textbf{\underline{N}}etwork (namely NFGNN). By estimating the node-oriented spectral filter for each node, NFGNN is provided with the capability of precise local node positioning via the generalized translated operator, thus discriminating the variations of local homophily patterns adaptively. Meanwhile, the utilization of re-parameterization brings a good trade-off between global consistency and local sensibility for learning the node-oriented spectral filters. Furthermore, we theoretically analyze the localization property of NFGNN, demonstrating that the signal after adaptive filtering is still positioned around the corresponding node. Extensive experimental results demonstrate that the proposed NFGNN achieves more favorable performance.

Partial and Asymmetric Contrastive Learning for Out-of-Distribution Detection in Long-Tailed Recognition

Existing out-of-distribution (OOD) detection methods are typically benchmarked on training sets with balanced class distributions. However, in real-world applications, it is common for the training sets to have long-tailed distributions. In this work, we first demonstrate that existing OOD detection methods commonly suffer from significant performance degradation when the training set is long-tail distributed. Through analysis, we posit that this is because the models struggle to distinguish the minority tail-class in-distribution samples, from the true OOD samples, making the tail classes more prone to be falsely detected as OOD. To solve this problem, we propose Partial and Asymmetric Supervised Contrastive Learning (PASCL), which explicitly encourages the model to distinguish between tail-class in-distribution samples and OOD samples. To further boost in-distribution classification accuracy, we propose Auxiliary Branch Finetuning, which uses two separate branches of BN and classification layers for anomaly detection and in-distribution classification, respectively. The intuition is that in-distribution and OOD anomaly data have different underlying distributions. Our method outperforms previous state-of-the-art method by $1.29\%$, $1.45\%$, $0.69\%$ anomaly detection false positive rate (FPR) and $3.24\%$, $4.06\%$, $7.89\%$ in-distribution classification accuracy on CIFAR10-LT, CIFAR100-LT, and ImageNet-LT, respectively. Code and pre-trained models are available at https://github.com/amazon-research/long-tailed-ood-detection.

Removing Batch Normalization Boosts Adversarial Training

Adversarial training (AT) defends deep neural networks against adversarial attacks. One challenge that limits its practical application is the performance degradation on clean samples. A major bottleneck identified by previous works is the widely used batch normalization (BN), which struggles to model the different statistics of clean and adversarial training samples in AT. Although the dominant approach is to extend BN to capture this mixture of distribution, we propose to completely eliminate this bottleneck by removing all BN layers in AT. Our normalizer-free robust training (NoFrost) method extends recent advances in normalizer-free networks to AT for its unexplored advantage on handling the mixture distribution challenge. We show that NoFrost achieves adversarial robustness with only a minor sacrifice on clean sample accuracy. On ImageNet with ResNet50, NoFrost achieves $74.06\%$ clean accuracy, which drops merely $2.00\%$ from standard training. In contrast, BN-based AT obtains $59.28\%$ clean accuracy, suffering a significant $16.78\%$ drop from standard training. In addition, NoFrost achieves a $23.56\%$ adversarial robustness against PGD attack, which improves the $13.57\%$ robustness in BN-based AT. We observe better model smoothness and larger decision margins from NoFrost, which make the models less sensitive to input perturbations and thus more robust. Moreover, when incorporating more data augmentations into NoFrost, it achieves comprehensive robustness against multiple distribution shifts. Code and pre-trained models are public at https://github.com/amazon-research/normalizer-free-robust-training.

Alexa Teacher Model: Pretraining and Distilling Multi-Billion-Parameter Encoders for Natural Language Understanding Systems

We present results from a large-scale experiment on pretraining encoders with non-embedding parameter counts ranging from 700M to 9.3B, their subsequent distillation into smaller models ranging from 17M-170M parameters, and their application to the Natural Language Understanding (NLU) component of a virtual assistant system. Though we train using 70% spoken-form data, our teacher models perform comparably to XLM-R and mT5 when evaluated on the written-form Cross-lingual Natural Language Inference (XNLI) corpus. We perform a second stage of pretraining on our teacher models using in-domain data from our system, improving error rates by 3.86% relative for intent classification and 7.01% relative for slot filling. We find that even a 170M-parameter model distilled from our Stage 2 teacher model has 2.88% better intent classification and 7.69% better slot filling error rates when compared to the 2.3B-parameter teacher trained only on public data (Stage 1), emphasizing the importance of in-domain data for pretraining. When evaluated offline using labeled NLU data, our 17M-parameter Stage 2 distilled model outperforms both XLM-R Base (85M params) and DistillBERT (42M params) by 4.23% to 6.14%, respectively. Finally, we present results from a full virtual assistant experimentation platform, where we find that models trained using our pretraining and distillation pipeline outperform models distilled from 85M-parameter teachers by 3.74%-4.91% on an automatic measurement of full-system user dissatisfaction.