Flowformer: Linearizing Transformers with Conservation Flows

Transformers based on the attention mechanism have achieved impressive success in various areas. However, the attention mechanism has a quadratic complexity, significantly impeding Transformers from dealing with numerous tokens and scaling up to bigger models. Previous methods mainly utilize the similarity decomposition and the associativity of matrix multiplication to devise linear-time attention mechanisms. They avoid degeneration of attention to a trivial distribution by reintroducing inductive biases such as the locality, thereby at the expense of model generality and expressiveness. In this paper, we linearize Transformers free from specific inductive biases based on the flow network theory. We cast attention as the information flow aggregated from the sources (values) to the sinks (results) through the learned flow capacities (attentions). Within this framework, we apply the property of flow conservation with attention and propose the Flow-Attention mechanism of linear complexity. By respectively conserving the incoming flow of sinks for source competition and the outgoing flow of sources for sink allocation, Flow-Attention inherently generates informative attentions without using specific inductive biases. Empowered by the Flow-Attention, Flowformer yields strong performance in linear time for wide areas, including long sequence, time series, vision, natural language, and reinforcement learning.

Supported Policy Optimization for Offline Reinforcement Learning

Policy constraint methods to offline reinforcement learning (RL) typically utilize parameterization or regularization that constrains the policy to perform actions within the support set of the behavior policy. The elaborative designs of parameterization methods usually intrude into the policy networks, which may bring extra inference cost and cannot take full advantage of well-established online methods. Regularization methods reduce the divergence between the learned policy and the behavior policy, which may mismatch the inherent density-based definition of support set thereby failing to avoid the out-of-distribution actions effectively. This paper presents Supported Policy OpTimization (SPOT), which is directly derived from the theoretical formalization of the density-based support constraint. SPOT adopts a VAE-based density estimator to explicitly model the support set of behavior policy and presents a simple but effective density-based regularization term, which can be plugged non-intrusively into off-the-shelf off-policy RL algorithms. On the standard benchmarks for offline RL, SPOT substantially outperforms state-of-the-art offline RL methods. Benefiting from the pluggable design, the offline pretrained models from SPOT can also be applied to perform online fine-tuning seamlessly.

Transferability in Deep Learning: A Survey

The success of deep learning algorithms generally depends on large-scale data, while humans appear to have inherent ability of knowledge transfer, by recognizing and applying relevant knowledge from previous learning experiences when encountering and solving unseen tasks. Such an ability to acquire and reuse knowledge is known as transferability in deep learning. It has formed the long-term quest towards making deep learning as data-efficient as human learning, and has been motivating fruitful design of more powerful deep learning algorithms. We present this survey to connect different isolated areas in deep learning with their relation to transferability, and to provide a unified and complete view to investigating transferability through the whole lifecycle of deep learning. The survey elaborates the fundamental goals and challenges in parallel with the core principles and methods, covering recent cornerstones in deep architectures, pre-training, task adaptation and domain adaptation. This highlights unanswered questions on the appropriate objectives for learning transferable knowledge and for adapting the knowledge to new tasks and domains, avoiding catastrophic forgetting and negative transfer. Finally, we implement a benchmark and an open-source library, enabling a fair evaluation of deep learning methods in terms of transferability.

Ranking and Tuning Pre-trained Models: A New Paradigm of Exploiting Model Hubs

Pre-trained model hubs with many pre-trained models (PTMs) have been a cornerstone in deep learning. Although built at a high cost, they are in fact \emph{under-exploited}: practitioners usually pick one PTM from the provided model hub by popularity, and then fine-tune the PTM to solve the target task. This na\"ve but common practice poses two obstacles to sufficiently exploiting pre-trained model hubs: (1) the PTM selection procedure has no optimality guarantee; (2) only one PTM is used while the rest PTMs are overlooked. Ideally, to maximally exploit pre-trained model hubs, trying all combinations of PTMs and extensively fine-tuning each combination of PTMs are required, which incurs exponential combinations and unaffordable computational budget. In this paper, we propose a new paradigm of exploiting model hubs by ranking and tuning pre-trained models: (1) Our conference work~\citep{you_logme:_2021} proposed LogME to estimate the maximum value of label evidence given features extracted by pre-trained models, which can rank all the PTMs in a model hub for various types of PTMs and tasks \emph{before fine-tuning}. (2) the best ranked PTM can be fine-tuned and deployed if we have no preference for the model's architecture, or the target PTM can be tuned by top-K ranked PTMs via the proposed B-Tuning algorithm. The ranking part is based on the conference paper, and we complete its theoretical analysis (convergence proof of the heuristic evidence maximization procedure, and the influence of feature dimension) in this paper. The tuning part introduces a novel Bayesian Tuning (B-Tuning) method for multiple PTMs tuning, which surpasses dedicated methods designed for homogeneous PTMs tuning and sets up new state of the art for heterogeneous PTMs tuning. We believe the new paradigm of exploiting PTM hubs can interest a large audience of the community.

Omni-Training for Data-Efficient Deep Learning

Learning a generalizable deep model from a few examples in a short time remains a major challenge of machine learning, which has impeded its wide deployment to many scenarios. Recent advances reveal that a properly pre-trained model endows an important property: transferability. A higher transferability of the learned representations indicates a better generalizability across domains of different distributions (domain transferability), or across tasks of different semantics (task transferability). Transferability has become the key to enable data-efficient deep learning, however, existing pre-training methods focus only on the domain transferability while meta-training methods only on the task transferability. This restricts their data-efficiency in downstream scenarios of diverging domains and tasks. A finding of this paper is that even a tight combination of pre-training and meta-training cannot achieve both kinds of transferability. This motivates the proposed Omni-Training framework towards data-efficient deep learning. Our first contribution is Omni-Net, a tri-flow architecture. Besides the joint representation flow, Omni-Net introduces two new parallel flows for pre-training and meta-training, respectively responsible for learning representations of domain transferability and task transferability. Omni-Net coordinates the parallel flows by routing them via the joint-flow, making each gain the other kind of transferability. Our second contribution is Omni-Loss, in which a mean-teacher regularization is imposed to learn generalizable and stabilized representations. Omni-Training is a general framework that accommodates many existing pre-training and meta-training algorithms. A thorough evaluation on cross-task and cross-domain datasets in classification, regression and reinforcement learning problems shows that Omni-Training consistently outperforms the state-of-the-art methods.

ModeRNN: Harnessing Spatiotemporal Mode Collapse in Unsupervised Predictive Learning

Learning predictive models for unlabeled spatiotemporal data is challenging in part because visual dynamics can be highly entangled in real scenes, making existing approaches prone to overfit partial modes of physical processes while neglecting to reason about others. We name this phenomenon spatiotemporal mode collapse and explore it for the first time in predictive learning. The key is to provide the model with a strong inductive bias to discover the compositional structures of latent modes. To this end, we propose ModeRNN, which introduces a novel method to learn structured hidden representations between recurrent states. The core idea of this framework is to first extract various components of visual dynamics using a set of spatiotemporal slots with independent parameters. Considering that multiple space-time patterns may co-exist in a sequence, we leverage learnable importance weights to adaptively aggregate slot features into a unified hidden representation, which is then used to update the recurrent states. Across the entire dataset, different modes result in different responses on the mixtures of slots, which enhances the ability of ModeRNN to build structured representations and thus prevents the so-called mode collapse. Unlike existing models, ModeRNN is shown to prevent spatiotemporal mode collapse and further benefit from learning mixed visual dynamics.

Anomaly Transformer: Time Series Anomaly Detection with Association Discrepancy

Unsupervisedly detecting anomaly points in time series is challenging, which requires the model to learn informative representations and derive a distinguishable criterion. Prior methods mainly detect anomalies based on the recurrent network representation of each time point. However, the point-wise representation is less informative for complex temporal patterns and can be dominated by normal patterns, making rare anomalies less distinguishable. We find that in each time series, each time point can also be described by its associations with all time points, presenting as a point-wise distribution that is more expressive for temporal modeling. We further observe that due to the rarity of anomalies, it is harder for anomalies to build strong associations with the whole series and their associations shall mainly concentrate on the adjacent time points. This observation implies an inherently distinguishable criterion between normal and abnormal points, which we highlight as the \emph{Association Discrepancy}. Technically we propose the \emph{Anomaly Transformer} with an \emph{Anomaly-Attention} mechanism to compute the association discrepancy. A minimax strategy is devised to amplify the normal-abnormal distinguishability of the association discrepancy. Anomaly Transformer achieves state-of-the-art performance on six unsupervised time series anomaly detection benchmarks for three applications: service monitoring, space \& earth exploration, and water treatment.

X-model: Improving Data Efficiency in Deep Learning with A Minimax Model

To mitigate the burden of data labeling, we aim at improving data efficiency for both classification and regression setups in deep learning. However, the current focus is on classification problems while rare attention has been paid to deep regression, which usually requires more human effort to labeling. Further, due to the intrinsic difference between categorical and continuous label space, the common intuitions for classification, e.g., cluster assumptions or pseudo labeling strategies, cannot be naturally adapted into deep regression. To this end, we first delved into the existing data-efficient methods in deep learning and found that they either encourage invariance to data stochasticity (e.g., consistency regularization under different augmentations) or model stochasticity (e.g., difference penalty for predictions of models with different dropout). To take the power of both worlds, we propose a novel X-model by simultaneously encouraging the invariance to {data stochasticity} and {model stochasticity}. Further, the X-model plays a minimax game between the feature extractor and task-specific heads to further enhance the invariance to model stochasticity. Extensive experiments verify the superiority of the X-model among various tasks, from a single-value prediction task of age estimation to a dense-value prediction task of keypoint localization, a 2D synthetic, and a 3D realistic dataset, as well as a multi-category object recognition task.

Decoupled Adaptation for Cross-Domain Object Detection

Cross-domain object detection is more challenging than object classification since multiple objects exist in an image and the location of each object is unknown in the unlabeled target domain. As a result, when we adapt features of different objects to enhance the transferability of the detector, the features of the foreground and the background are easy to be confused, which may hurt the discriminability of the detector. Besides, previous methods focused on category adaptation but ignored another important part for object detection, i.e., the adaptation on bounding box regression. To this end, we propose D-adapt, namely Decoupled Adaptation, to decouple the adversarial adaptation and the training of the detector. Besides, we fill the blank of regression domain adaptation in object detection by introducing a bounding box adaptor. Experiments show that D-adapt achieves state-of-the-art results on four cross-domain object detection tasks and yields 17% and 21% relative improvement on benchmark datasets Clipart1k and Comic2k in particular.

Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting

Extending the forecasting time is a critical demand for real applications, such as extreme weather early warning and long-term energy consumption planning. This paper studies the \textit{long-term forecasting} problem of time series. Prior Transformer-based models adopt various self-attention mechanisms to discover the long-range dependencies. However, intricate temporal patterns of the long-term future prohibit the model from finding reliable dependencies. Also, Transformers have to adopt the sparse versions of point-wise self-attentions for long series efficiency, resulting in the information utilization bottleneck. Towards these challenges, we propose Autoformer as a novel decomposition architecture with an Auto-Correlation mechanism. We go beyond the pre-processing convention of series decomposition and renovate it as a basic inner block of deep models. This design empowers Autoformer with progressive decomposition capacities for complex time series. Further, inspired by the stochastic process theory, we design the Auto-Correlation mechanism based on the series periodicity, which conducts the dependencies discovery and representation aggregation at the sub-series level. Auto-Correlation outperforms self-attention in both efficiency and accuracy. In long-term forecasting, Autoformer yields state-of-the-art accuracy, with a 38% relative improvement on six benchmarks, covering five practical applications: energy, traffic, economics, weather and disease.