Toward Efficient Language Model Pretraining and Downstream Adaptation via Self-Evolution: A Case Study on SuperGLUE

This technical report briefly describes our JDExplore d-team's Vega v2 submission on the SuperGLUE leaderboard. SuperGLUE is more challenging than the widely used general language understanding evaluation (GLUE) benchmark, containing eight difficult language understanding tasks, including question answering, natural language inference, word sense disambiguation, coreference resolution, and reasoning. [Method] Instead of arbitrarily increasing the size of a pretrained language model (PLM), our aim is to 1) fully extract knowledge from the input pretraining data given a certain parameter budget, e.g., 6B, and 2) effectively transfer this knowledge to downstream tasks. To achieve goal 1), we propose self-evolution learning for PLMs to wisely predict the informative tokens that should be masked, and supervise the masked language modeling (MLM) process with rectified smooth labels. For goal 2), we leverage the prompt transfer technique to improve the low-resource tasks by transferring the knowledge from the foundation model and related downstream tasks to the target task. [Results] According to our submission record (Oct. 2022), with our optimized pretraining and fine-tuning strategies, our 6B Vega method achieved new state-of-the-art performance on 4/8 tasks, sitting atop the SuperGLUE leaderboard on Oct. 8, 2022, with an average score of 91.3.

Knowledge-Aware Federated Active Learning with Non-IID Data

Federated learning enables multiple decentralized clients to learn collaboratively without sharing the local training data. However, the expensive annotation cost to acquire data labels on local clients remains an obstacle in utilizing local data. In this paper, we propose a federated active learning paradigm to efficiently learn a global model with limited annotation budget while protecting data privacy in a decentralized learning way. The main challenge faced by federated active learning is the mismatch between the active sampling goal of the global model on the server and that of the asynchronous local clients. This becomes even more significant when data is distributed non-IID across local clients. To address the aforementioned challenge, we propose Knowledge-Aware Federated Active Learning (KAFAL), which consists of Knowledge-Specialized Active Sampling (KSAS) and Knowledge-Compensatory Federated Update (KCFU). KSAS is a novel active sampling method tailored for the federated active learning problem. It deals with the mismatch challenge by sampling actively based on the discrepancies between local and global models. KSAS intensifies specialized knowledge in local clients, ensuring the sampled data to be informative for both the local clients and the global model. KCFU, in the meantime, deals with the client heterogeneity caused by limited data and non-IID data distributions. It compensates for each client's ability in weak classes by the assistance of the global model. Extensive experiments and analyses are conducted to show the superiority of KSAS over the state-of-the-art active learning methods and the efficiency of KCFU under the federated active learning framework.

Responsible Active Learning via Human-in-the-loop Peer Study

Active learning has been proposed to reduce data annotation efforts by only manually labelling representative data samples for training. Meanwhile, recent active learning applications have benefited a lot from cloud computing services with not only sufficient computational resources but also crowdsourcing frameworks that include many humans in the active learning loop. However, previous active learning methods that always require passing large-scale unlabelled data to cloud may potentially raise significant data privacy issues. To mitigate such a risk, we propose a responsible active learning method, namely Peer Study Learning (PSL), to simultaneously preserve data privacy and improve model stability. Specifically, we first introduce a human-in-the-loop teacher-student architecture to isolate unlabelled data from the task learner (teacher) on the cloud-side by maintaining an active learner (student) on the client-side. During training, the task learner instructs the light-weight active learner which then provides feedback on the active sampling criterion. To further enhance the active learner via large-scale unlabelled data, we introduce multiple peer students into the active learner which is trained by a novel learning paradigm, including the In-Class Peer Study on labelled data and the Out-of-Class Peer Study on unlabelled data. Lastly, we devise a discrepancy-based active sampling criterion, Peer Study Feedback, that exploits the variability of peer students to select the most informative data to improve model stability. Extensive experiments demonstrate the superiority of the proposed PSL over a wide range of active learning methods in both standard and sensitive protection settings.

Domain-Specific Risk Minimization

Learning a domain-invariant representation has become one of the most popular approaches for domain adaptation/generalization. In this paper, we show that the invariant representation may not be sufficient to guarantee a good generalization, where the labeling function shift should be taken into consideration. Inspired by this, we first derive a new generalization upper bound on the empirical risk that explicitly considers the labeling function shift. We then propose Domain-specific Risk Minimization (DRM), which can model the distribution shifts of different domains separately and select the most appropriate one for the target domain. Extensive experiments on four popular domain generalization datasets, CMNIST, PACS, VLCS, and DomainNet, demonstrate the effectiveness of the proposed DRM for domain generalization with the following advantages: 1) it significantly outperforms competitive baselines; 2) it enables either comparable or superior accuracies on all training domains comparing to vanilla empirical risk minimization (ERM); 3) it remains very simple and efficient during training, and 4) it is complementary to invariant learning approaches.

Improving Fine-Grained Visual Recognition in Low Data Regimes via Self-Boosting Attention Mechanism

The challenge of fine-grained visual recognition often lies in discovering the key discriminative regions. While such regions can be automatically identified from a large-scale labeled dataset, a similar method might become less effective when only a few annotations are available. In low data regimes, a network often struggles to choose the correct regions for recognition and tends to overfit spurious correlated patterns from the training data. To tackle this issue, this paper proposes the self-boosting attention mechanism, a novel method for regularizing the network to focus on the key regions shared across samples and classes. Specifically, the proposed method first generates an attention map for each training image, highlighting the discriminative part for identifying the ground-truth object category. Then the generated attention maps are used as pseudo-annotations. The network is enforced to fit them as an auxiliary task. We call this approach the self-boosting attention mechanism (SAM). We also develop a variant by using SAM to create multiple attention maps to pool convolutional maps in a style of bilinear pooling, dubbed SAM-Bilinear. Through extensive experimental studies, we show that both methods can significantly improve fine-grained visual recognition performance on low data regimes and can be incorporated into existing network architectures. The source code is publicly available at: https://github.com/GANPerf/SAM

MeshMAE: Masked Autoencoders for 3D Mesh Data Analysis

Recently, self-supervised pre-training has advanced Vision Transformers on various tasks w.r.t. different data modalities, e.g., image and 3D point cloud data. In this paper, we explore this learning paradigm for 3D mesh data analysis based on Transformers. Since applying Transformer architectures to new modalities is usually non-trivial, we first adapt Vision Transformer to 3D mesh data processing, i.e., Mesh Transformer. In specific, we divide a mesh into several non-overlapping local patches with each containing the same number of faces and use the 3D position of each patch's center point to form positional embeddings. Inspired by MAE, we explore how pre-training on 3D mesh data with the Transformer-based structure benefits downstream 3D mesh analysis tasks. We first randomly mask some patches of the mesh and feed the corrupted mesh into Mesh Transformers. Then, through reconstructing the information of masked patches, the network is capable of learning discriminative representations for mesh data. Therefore, we name our method MeshMAE, which can yield state-of-the-art or comparable performance on mesh analysis tasks, i.e., classification and segmentation. In addition, we also conduct comprehensive ablation studies to show the effectiveness of key designs in our method.

Deep Dictionary Learning with An Intra-class Constraint

In recent years, deep dictionary learning (DDL)has attracted a great amount of attention due to its effectiveness for representation learning and visual recognition.~However, most existing methods focus on unsupervised deep dictionary learning, failing to further explore the category information.~To make full use of the category information of different samples, we propose a novel deep dictionary learning model with an intra-class constraint (DDLIC) for visual classification. Specifically, we design the intra-class compactness constraint on the intermediate representation at different levels to encourage the intra-class representations to be closer to each other, and eventually the learned representation becomes more discriminative.~Unlike the traditional DDL methods, during the classification stage, our DDLIC performs a layer-wise greedy optimization in a similar way to the training stage. Experimental results on four image datasets show that our method is superior to the state-of-the-art methods.

GFNet: Geometric Flow Network for 3D Point Cloud Semantic Segmentation

Point cloud semantic segmentation from projected views, such as range-view (RV) and bird's-eye-view (BEV), has been intensively investigated. Different views capture different information of point clouds and thus are complementary to each other. However, recent projection-based methods for point cloud semantic segmentation usually utilize a vanilla late fusion strategy for the predictions of different views, failing to explore the complementary information from a geometric perspective during the representation learning. In this paper, we introduce a geometric flow network (GFNet) to explore the geometric correspondence between different views in an align-before-fuse manner. Specifically, we devise a novel geometric flow module (GFM) to bidirectionally align and propagate the complementary information across different views according to geometric relationships under the end-to-end learning scheme. We perform extensive experiments on two widely used benchmark datasets, SemanticKITTI and nuScenes, to demonstrate the effectiveness of our GFNet for project-based point cloud semantic segmentation. Concretely, GFNet not only significantly boosts the performance of each individual view but also achieves state-of-the-art results over all existing projection-based models. Code is available at \url{https://github.com/haibo-qiu/GFNet}.

BatchFormerV2: Exploring Sample Relationships for Dense Representation Learning

Attention mechanisms have been very popular in deep neural networks, where the Transformer architecture has achieved great success in not only natural language processing but also visual recognition applications. Recently, a new Transformer module, applying on batch dimension rather than spatial/channel dimension, i.e., BatchFormer [18], has been introduced to explore sample relationships for overcoming data scarcity challenges. However, it only works with image-level representations for classification. In this paper, we devise a more general batch Transformer module, BatchFormerV2, which further enables exploring sample relationships for dense representation learning. Specifically, when applying the proposed module, it employs a two-stream pipeline during training, i.e., either with or without a BatchFormerV2 module, where the batchformer stream can be removed for testing. Therefore, the proposed method is a plug-and-play module and can be easily integrated into different vision Transformers without any extra inference cost. Without bells and whistles, we show the effectiveness of the proposed method for a variety of popular visual recognition tasks, including image classification and two important dense prediction tasks: object detection and panoptic segmentation. Particularly, BatchFormerV2 consistently improves current DETR-based detection methods (e.g., DETR, Deformable-DETR, Conditional DETR, and SMCA) by over 1.3%. Code will be made publicly available.

BatchFormer: Learning to Explore Sample Relationships for Robust Representation Learning

Despite the success of deep neural networks, there are still many challenges in deep representation learning due to the data scarcity issues such as data imbalance, unseen distribution, and domain shift. To address the above-mentioned issues, a variety of methods have been devised to explore the sample relationships in a vanilla way (i.e., from the perspectives of either the input or the loss function), failing to explore the internal structure of deep neural networks for learning with sample relationships. Inspired by this, we propose to enable deep neural networks themselves with the ability to learn the sample relationships from each mini-batch. Specifically, we introduce a batch transformer module or BatchFormer, which is then applied into the batch dimension of each mini-batch to implicitly explore sample relationships during training. By doing this, the proposed method enables the collaboration of different samples, e.g., the head-class samples can also contribute to the learning of the tail classes for long-tailed recognition. Furthermore, to mitigate the gap between training and testing, we share the classifier between with or without the BatchFormer during training, which can thus be removed during testing. We perform extensive experiments on over ten datasets and the proposed method achieves significant improvements on different data scarcity applications without any bells and whistles, including the tasks of long-tailed recognition, compositional zero-shot learning, domain generalization, and contrastive learning. Code will be made publicly available at https://github.com/zhihou7/BatchFormer.