Token merging has emerged as a new paradigm that can accelerate the inference of Vision Transformers (ViTs) without any retraining or fine-tuning. To push the frontier of training-free acceleration in ViTs, we improve token merging by adding the perspectives of 1) activation outliers and 2) hierarchical representations. Through a careful analysis of the attention behavior in ViTs, we characterize a delayed onset of the convergent attention phenomenon, which makes token merging undesirable in the bottom blocks of ViTs. Moreover, we augment token merging with a hierarchical processing scheme to capture multi-scale redundancy between visual tokens. Combining these two insights, we build a unified inference framework called DSM: Delayed Spatial Merging. We extensively evaluate DSM on various ViT model scales (Tiny to Huge) and tasks (ImageNet-1k and transfer learning), achieving up to 1.8$\times$ FLOP reduction and 1.6$\times$ throughput speedup at a negligible loss while being two orders of magnitude faster than existing methods.
Recently, data-driven models such as deep neural networks have shown to be promising tools for modelling and state inference in soft robots. However, voluminous amounts of data are necessary for deep models to perform effectively, which requires exhaustive and quality data collection, particularly of state labels. Consequently, obtaining labelled state data for soft robotic systems is challenged for various reasons, including difficulty in the sensorization of soft robots and the inconvenience of collecting data in unstructured environments. To address this challenge, in this paper, we propose a semi-supervised sequential variational Bayes (DSVB) framework for transfer learning and state inference in soft robots with missing state labels on certain robot configurations. Considering that soft robots may exhibit distinct dynamics under different robot configurations, a feature space transfer strategy is also incorporated to promote the adaptation of latent features across multiple configurations. Unlike existing transfer learning approaches, our proposed DSVB employs a recurrent neural network to model the nonlinear dynamics and temporal coherence in soft robot data. The proposed framework is validated on multiple setup configurations of a pneumatic-based soft robot finger. Experimental results on four transfer scenarios demonstrate that DSVB performs effective transfer learning and accurate state inference amidst missing state labels. The data and code are available at https://github.com/shageenderan/DSVB.
Graph Neural Networks (GNNs) have become powerful tools in modeling graph-structured data in recommender systems. However, real-life recommendation scenarios usually involve heterogeneous relationships (e.g., social-aware user influence, knowledge-aware item dependency) which contains fruitful information to enhance the user preference learning. In this paper, we study the problem of heterogeneous graph-enhanced relational learning for recommendation. Recently, contrastive self-supervised learning has become successful in recommendation. In light of this, we propose a Heterogeneous Graph Contrastive Learning (HGCL), which is able to incorporate heterogeneous relational semantics into the user-item interaction modeling with contrastive learning-enhanced knowledge transfer across different views. However, the influence of heterogeneous side information on interactions may vary by users and items. To move this idea forward, we enhance our heterogeneous graph contrastive learning with meta networks to allow the personalized knowledge transformer with adaptive contrastive augmentation. The experimental results on three real-world datasets demonstrate the superiority of HGCL over state-of-the-art recommendation methods. Through ablation study, key components in HGCL method are validated to benefit the recommendation performance improvement. The source code of the model implementation is available at the link https://github.com/HKUDS/HGCL.
Transfer learning has remarkably improved computer vision. These advances also promise improvements in neuroimaging, where training set sizes are often small. However, various difficulties arise in directly applying models pretrained on natural images to radiologic images, such as MRIs. In particular, a mismatch in the input space (2D images vs. 3D MRIs) restricts the direct transfer of models, often forcing us to consider only a few MRI slices as input. To this end, we leverage the 2D-Slice-CNN architecture of Gupta et al. (2021), which embeds all the MRI slices with 2D encoders (neural networks that take 2D image input) and combines them via permutation-invariant layers. With the insight that the pretrained model can serve as the 2D encoder, we initialize the 2D encoder with ImageNet pretrained weights that outperform those initialized and trained from scratch on two neuroimaging tasks -- brain age prediction on the UK Biobank dataset and Alzheimer's disease detection on the ADNI dataset. Further, we improve the modeling capabilities of 2D-Slice models by incorporating spatial information through position embeddings, which can improve the performance in some cases.
Recent recommender systems have shown remarkable performance by using an ensemble of heterogeneous models. However, it is exceedingly costly because it requires resources and inference latency proportional to the number of models, which remains the bottleneck for production. Our work aims to transfer the ensemble knowledge of heterogeneous teachers to a lightweight student model using knowledge distillation (KD), to reduce the huge inference costs while retaining high accuracy. Through an empirical study, we find that the efficacy of distillation severely drops when transferring knowledge from heterogeneous teachers. Nevertheless, we show that an important signal to ease the difficulty can be obtained from the teacher's training trajectory. This paper proposes a new KD framework, named HetComp, that guides the student model by transferring easy-to-hard sequences of knowledge generated from the teachers' trajectories. To provide guidance according to the student's learning state, HetComp uses dynamic knowledge construction to provide progressively difficult ranking knowledge and adaptive knowledge transfer to gradually transfer finer-grained ranking information. Our comprehensive experiments show that HetComp significantly improves the distillation quality and the generalization of the student model.
To reap the promising benefits of massive multiple-input multiple-output (MIMO) systems, accurate channel state information (CSI) is required through channel estimation. However, due to the complicated wireless propagation environment and large-scale antenna arrays, precise channel estimation for massive MIMO systems is significantly challenging and costs an enormous training overhead. Considerable time-frequency resources are consumed to acquire sufficient accuracy of CSI, which thus severely degrades systems' spectral and energy efficiencies. In this paper, we propose a dual-attention-based channel estimation network (DACEN) to realize accurate channel estimation via low-density pilots, by jointly learning the spatial-temporal domain features of massive MIMO channels with the temporal attention module and the spatial attention module. To further improve the estimation accuracy, we propose a parameter-instance transfer learning approach to transfer the channel knowledge learned from the high-density pilots pre-acquired during the training dataset collection period. Experimental results reveal that the proposed DACEN-based method achieves better channel estimation performance than the existing methods under various pilot-density settings and signal-to-noise ratios. Additionally, with the proposed parameter-instance transfer learning approach, the DACEN-based method achieves additional performance gain, thereby further demonstrating the effectiveness and superiority of the proposed method.
Physics-informed neural networks (PINNs) as a means of solving partial differential equations (PDE) have garnered much attention in the Computational Science and Engineering (CS&E) world. However, a recent topic of interest is exploring various training (i.e., optimization) challenges - in particular, arriving at poor local minima in the optimization landscape results in a PINN approximation giving an inferior, and sometimes trivial, solution when solving forward time-dependent PDEs with no data. This problem is also found in, and in some sense more difficult, with domain decomposition strategies such as temporal decomposition using XPINNs. We furnish examples and explanations for different training challenges, their cause, and how they relate to information propagation and temporal decomposition. We then propose a new stacked-decomposition method that bridges the gap between time-marching PINNs and XPINNs. We also introduce significant computational speed-ups by using transfer learning concepts to initialize subnetworks in the domain and loss tolerance-based propagation for the subdomains. Finally, we formulate a new time-sweeping collocation point algorithm inspired by the previous PINNs causality literature, which our framework can still describe, and provides a significant computational speed-up via reduced-cost collocation point segmentation. The proposed methods form our unified framework, which overcomes training challenges in PINNs and XPINNs for time-dependent PDEs by respecting the causality in multiple forms and improving scalability by limiting the computation required per optimization iteration. Finally, we provide numerical results for these methods on baseline PDE problems for which unmodified PINNs and XPINNs struggle to train.
Link prediction is central to many real-world applications, but its performance may be hampered when the graph of interest is sparse. To alleviate issues caused by sparsity, we investigate a previously overlooked phenomenon: in many cases, a densely connected, complementary graph can be found for the original graph. The denser graph may share nodes with the original graph, which offers a natural bridge for transferring selective, meaningful knowledge. We identify this setting as Graph Intersection-induced Transfer Learning (GITL), which is motivated by practical applications in e-commerce or academic co-authorship predictions. We develop a framework to effectively leverage the structural prior in this setting. We first create an intersection subgraph using the shared nodes between the two graphs, then transfer knowledge from the source-enriched intersection subgraph to the full target graph. In the second step, we consider two approaches: a modified label propagation, and a multi-layer perceptron (MLP) model in a teacher-student regime. Experimental results on proprietary e-commerce datasets and open-source citation graphs show that the proposed workflow outperforms existing transfer learning baselines that do not explicitly utilize the intersection structure.
The cultural heritage buildings (CHB), which are part of mankind's history and identity, are in constant danger of damage or in extreme situations total destruction. That being said, it's of utmost importance to preserve them by identifying the existent, or presumptive, defects using novel methods so that renovation processes can be done in a timely manner and with higher accuracy. The main goal of this research is to use new deep learning (DL) methods in the process of preserving CHBs (situated in Iran); a goal that has been neglected especially in developing countries such as Iran, as these countries still preserve their CHBs using manual, and even archaic, methods that need direct human supervision. Having proven their effectiveness and performance when it comes to processing images, the convolutional neural networks (CNN) are a staple in computer vision (CV) literacy and this paper is not exempt. When lacking enough CHB images, training a CNN from scratch would be very difficult and prone to overfitting; that's why we opted to use a technique called transfer learning (TL) in which we used pre-trained ResNet, MobileNet, and Inception networks, for classification. Even more, the Grad-CAM was utilized to localize the defects to some extent. The final results were very favorable based on those of similar research. The final proposed model can pave the way for moving from manual to unmanned CHB conservation, hence an increase in accuracy and a decrease in human-induced errors.
We consider the continual representation learning setting: sequentially pretrain a model $M'$ on tasks $T_1, \ldots, T_T$, and then adapt $M'$ on a small amount of data from each task $T_i$ to check if it has forgotten information from old tasks. Under a kNN adaptation protocol, prior work shows that continual learning methods improve forgetting over naive training (SGD). In reality, practitioners do not use kNN classifiers -- they use the adaptation method that works best (e.g., fine-tuning) -- here, we find that strong continual learning baselines do worse than naive training. Interestingly, we find that a method from the transfer learning community (LP-FT) outperforms naive training and the other continual learning methods. Even with standard kNN evaluation protocols, LP-FT performs comparably with strong continual learning methods (while being simpler and requiring less memory) on three standard benchmarks: sequential CIFAR-10, CIFAR-100, and TinyImageNet. LP-FT also reduces forgetting in a real world satellite remote sensing dataset (FMoW), and a variant of LP-FT gets state-of-the-art accuracies on an NLP continual learning benchmark.
With the demanding need for deploying dialogue systems in new domains with less cost, zero-shot dialogue state tracking (DST), which tracks user's requirements in task-oriented dialogues without training on desired domains, draws attention increasingly. Although prior works have leveraged question-answering (QA) data to reduce the need for in-domain training in DST, they fail to explicitly model knowledge transfer and fusion for tracking dialogue states. To address this issue, we propose CoFunDST, which is trained on domain-agnostic QA datasets and directly uses candidate choices of slot-values as knowledge for zero-shot dialogue-state generation, based on a T5 pre-trained language model. Specifically, CoFunDST selects highly-relevant choices to the reference context and fuses them to initialize the decoder to constrain the model outputs. Our experimental results show that our proposed model achieves outperformed joint goal accuracy compared to existing zero-shot DST approaches in most domains on the MultiWOZ 2.1. Extensive analyses demonstrate the effectiveness of our proposed approach for improving zero-shot DST learning from QA.
The transfer learning paradigm of model pre-training and subsequent fine-tuning produces high-accuracy models. While most studies recommend scaling the pre-training size to benefit most from transfer learning, a question remains: what data and method should be used for pre-training? We investigate the impact of pre-training data distribution on the few-shot and full fine-tuning performance using 3 pre-training methods (supervised, contrastive language-image and image-image), 7 pre-training datasets, and 9 downstream datasets. Through extensive controlled experiments, we find that the choice of the pre-training data source is essential for the few-shot transfer, but its role decreases as more data is made available for fine-tuning. Additionally, we explore the role of data curation and examine the trade-offs between label noise and the size of the pre-training dataset. We find that using 2000X more pre-training data from LAION can match the performance of supervised ImageNet pre-training. Furthermore, we investigate the effect of pre-training methods, comparing language-image contrastive vs. image-image contrastive, and find that the latter leads to better downstream accuracy
Speech models have long been known to overfit individual speakers for many classification tasks. This leads to poor generalization in settings where the speakers are out-of-domain or out-of-distribution, as is common in production environments. We view speaker adaptation as a few-shot learning problem and propose investigating transfer learning approaches inspired by recent success with pre-trained models in natural language tasks. We propose pre-finetuning speech models on difficult tasks to distill knowledge into few-shot downstream classification objectives. We pre-finetune Wav2Vec2.0 on every permutation of four multiclass emotional speech recognition corpora and evaluate our pre-finetuned models through 33,600 few-shot fine-tuning trials on the Emotional Speech Dataset.
We propose a tuning-free dynamic SGD step size formula, which we call Distance over Gradients (DoG). The DoG step sizes depend on simple empirical quantities (distance from the initial point and norms of gradients) and have no ``learning rate'' parameter. Theoretically, we show that a slight variation of the DoG formula enjoys strong parameter-free convergence guarantees for stochastic convex optimization assuming only \emph{locally bounded} stochastic gradients. Empirically, we consider a broad range of vision and language transfer learning tasks, and show that DoG's performance is close to that of SGD with tuned learning rate. We also propose a per-layer variant of DoG that generally outperforms tuned SGD, approaching the performance of tuned Adam. A PyTorch implementation is available at https://github.com/formll/dog
Accurate traffic prediction is essential for effective urban management and the improvement of transportation efficiency. Recently, data-driven traffic prediction methods have been widely adopted, with better performance than traditional approaches. However, they often require large amounts of data for effective training, which becomes challenging given the prevalence of data scarcity in regions with inadequate sensing infrastructures. To address this issue, we propose a Semantic-Fused Multi-Granularity Transfer Learning (SFMGTL) model to achieve knowledge transfer across cities with fused semantics at different granularities. In detail, we design a semantic fusion module to fuse various semantics while conserving static spatial dependencies via reconstruction losses. Then, a fused graph is constructed based on node features through graph structure learning. Afterwards, we implement hierarchical node clustering to generate graphs with different granularity. To extract feasible meta-knowledge, we further introduce common and private memories and obtain domain-invariant features via adversarial training. It is worth noting that our work jointly addresses semantic fusion and multi-granularity issues in transfer learning. We conduct extensive experiments on six real-world datasets to verify the effectiveness of our SFMGTL model by comparing it with other state-of-the-art baselines. Afterwards, we also perform ablation and case studies, demonstrating that our model possesses substantially fewer parameters compared to baseline models. Moreover, we illustrate how knowledge transfer aids the model in accurately predicting demands, especially during peak hours. The codes can be found at https://github.com/zeonchen/SFMGTL.
Image-to-image translation is a fundamental task in computer vision. It transforms images from one domain to images in another domain so that they have particular domain-specific characteristics. Most prior works train a generative model to learn the mapping from a source domain to a target domain. However, learning such mapping between domains is challenging because data from different domains can be highly unbalanced in terms of both quality and quantity. To address this problem, we propose a new approach to extract image features by learning the similarities and differences of samples within the same data distribution via a novel contrastive learning framework, which we call Auto-Contrastive-Encoder (ACE). ACE learns the content code as the similarity between samples with the same content information and different style perturbations. The design of ACE enables us to achieve zero-shot image-to-image translation with no training on image translation tasks for the first time. Moreover, our learning method can learn the style features of images on different domains effectively. Consequently, our model achieves competitive results on multimodal image translation tasks with zero-shot learning as well. Additionally, we demonstrate the potential of our method in transfer learning. With fine-tuning, the quality of translated images improves in unseen domains. Even though we use contrastive learning, all of our training can be performed on a single GPU with the batch size of 8.
Scaled dot-product attention applies a softmax function on the scaled dot-product of queries and keys to calculate weights and then multiplies the weights and values. In this work, we study how to improve the learning of scaled dot-product attention to improve the accuracy of DETR. Our method is based on the following observations: using ground truth foreground-background mask (GT Fg-Bg Mask) as additional cues in the weights/values learning enables learning much better weights/values; with better weights/values, better values/weights can be learned. We propose a triple-attention module in which the first attention is a plain scaled dot-product attention, the second/third attention generates high-quality weights/values (with the assistance of GT Fg-Bg Mask) and shares the values/weights with the first attention to improve the quality of values/weights. The second and third attentions are removed during inference. We call our method knowledge-sharing DETR (KS-DETR), which is an extension of knowledge distillation (KD) in the way that the improved weights and values of the teachers (the second and third attentions) are directly shared, instead of mimicked, by the student (the first attention) to enable more efficient knowledge transfer from the teachers to the student. Experiments on various DETR-like methods show consistent improvements over the baseline methods on the MS COCO benchmark. Code is available at https://github.com/edocanonymous/KS-DETR.
Backdoor data detection is traditionally studied in an end-to-end supervised learning (SL) setting. However, recent years have seen the proliferating adoption of self-supervised learning (SSL) and transfer learning (TL), due to their lesser need for labeled data. Successful backdoor attacks have also been demonstrated in these new settings. However, we lack a thorough understanding of the applicability of existing detection methods across a variety of learning settings. By evaluating 56 attack settings, we show that the performance of most existing detection methods varies significantly across different attacks and poison ratios, and all fail on the state-of-the-art clean-label attack. In addition, they either become inapplicable or suffer large performance losses when applied to SSL and TL. We propose a new detection method called Active Separation via Offset (ASSET), which actively induces different model behaviors between the backdoor and clean samples to promote their separation. We also provide procedures to adaptively select the number of suspicious points to remove. In the end-to-end SL setting, ASSET is superior to existing methods in terms of consistency of defensive performance across different attacks and robustness to changes in poison ratios; in particular, it is the only method that can detect the state-of-the-art clean-label attack. Moreover, ASSET's average detection rates are higher than the best existing methods in SSL and TL, respectively, by 69.3% and 33.2%, thus providing the first practical backdoor defense for these new DL settings. We open-source the project to drive further development and encourage engagement: https://github.com/ruoxi-jia-group/ASSET.
Recent advances in training vision-language models have demonstrated unprecedented robustness and transfer learning effectiveness; however, standard computer vision datasets are image-only, and therefore not well adapted to such training methods. Our paper introduces a simple methodology for adapting any fine-grained image classification dataset for distributed vision-language pretraining. We implement this methodology on the challenging iNaturalist-2021 dataset, comprised of approximately 2.7 million images of macro-organisms across 10,000 classes, and achieve a new state-of-the art model in terms of zero-shot classification accuracy. Somewhat surprisingly, our model (trained using a new method called locked-image text tuning) uses a pre-trained, frozen vision representation, proving that language alignment alone can attain strong transfer learning performance, even on fractious, long-tailed datasets. Our approach opens the door for utilizing high quality vision-language pretrained models in agriculturally relevant applications involving species detection.
Amyotrophic Lateral Sclerosis (ALS) is a complex neurodegenerative disorder involving motor neuron degeneration. Significant research has begun to establish brain magnetic resonance imaging (MRI) as a potential biomarker to diagnose and monitor the state of the disease. Deep learning has turned into a prominent class of machine learning programs in computer vision and has been successfully employed to solve diverse medical image analysis tasks. However, deep learning-based methods applied to neuroimaging have not achieved superior performance in ALS patients classification from healthy controls due to having insignificant structural changes correlated with pathological features. Therefore, the critical challenge in deep models is to determine useful discriminative features with limited training data. By exploiting the long-range relationship of image features, this study introduces a framework named SF2Former that leverages vision transformer architecture's power to distinguish the ALS subjects from the control group. To further improve the network's performance, spatial and frequency domain information are combined because MRI scans are captured in the frequency domain before being converted to the spatial domain. The proposed framework is trained with a set of consecutive coronal 2D slices, which uses the pre-trained weights on ImageNet by leveraging transfer learning. Finally, a majority voting scheme has been employed to those coronal slices of a particular subject to produce the final classification decision. Our proposed architecture has been thoroughly assessed with multi-modal neuroimaging data using two well-organized versions of the Canadian ALS Neuroimaging Consortium (CALSNIC) multi-center datasets. The experimental results demonstrate the superiority of our proposed strategy in terms of classification accuracy compared with several popular deep learning-based techniques.