The gap between speech and text modalities is a major challenge in speech-to-text translation (ST). Different methods have been proposed for reducing this gap, but most of them require architectural changes in ST training. In this work, we propose to mitigate this issue at the pre-training stage, requiring no change in the ST model. First, we show that the connectionist temporal classification (CTC) loss can reduce the modality gap by design. We provide a quantitative comparison with the more common cross-entropy loss, showing that pre-training with CTC consistently achieves better final ST accuracy. Nevertheless, CTC is only a partial solution and thus, in our second contribution, we propose a novel pre-training method combining CTC and optimal transport to further reduce this gap. Our method pre-trains a Siamese-like model composed of two encoders, one for acoustic inputs and the other for textual inputs, such that they produce representations that are close to each other in the Wasserstein space. Extensive experiments on the standard CoVoST-2 and MuST-C datasets show that our pre-training method applied to the vanilla encoder-decoder Transformer achieves state-of-the-art performance under the no-external-data setting, and performs on par with recent strong multi-task learning systems trained with external data. Finally, our method can also be applied on top of these multi-task systems, leading to further improvements for these models.
相關內容
We study the problem of semantic segmentation calibration. Lots of solutions have been proposed to approach model miscalibration of confidence in image classification. However, to date, confidence calibration research on semantic segmentation is still limited. We provide a systematic study on the calibration of semantic segmentation models and propose a simple yet effective approach. First, we find that model capacity, crop size, multi-scale testing, and prediction correctness have impact on calibration. Among them, prediction correctness, especially misprediction, is more important to miscalibration due to over-confidence. Next, we propose a simple, unifying, and effective approach, namely selective scaling, by separating correct/incorrect prediction for scaling and more focusing on misprediction logit smoothing. Then, we study popular existing calibration methods and compare them with selective scaling on semantic segmentation calibration. We conduct extensive experiments with a variety of benchmarks on both in-domain and domain-shift calibration, and show that selective scaling consistently outperforms other methods.
Finding meaningful ways to measure the statistical dependency between random variables $\xi$ and $\zeta$ is a timeless statistical endeavor. In recent years, several novel concepts, like the distance covariance, have extended classical notions of dependency to more general settings. In this article, we propose and study an alternative framework that is based on optimal transport. The transport dependency $\tau \ge 0$ applies to general Polish spaces and intrinsically respects metric properties. For suitable ground costs, independence is fully characterized by $\tau = 0$. Via proper normalization of $\tau$, three transport correlations $\rho_\alpha$, $\rho_\infty$, and $\rho_*$ with values in $[0, 1]$ are defined. They attain the value $1$ if and only if $\zeta = \varphi(\xi)$, where $\varphi$ is an $\alpha$-Lipschitz function for $\rho_\alpha$, a measurable function for $\rho_\infty$, or a multiple of an isometry for $\rho_*$. The transport dependency can be estimated consistently by an empirical plug-in approach, but alternative estimators with the same convergence rate but significantly reduced computational costs are also proposed. Numerical results suggest that $\tau$ robustly recovers dependency between data sets with different internal metric structures. The usage for inferential tasks, like transport dependency based independence testing, is illustrated on a data set from a cancer study.
Autoencoders have demonstrated remarkable success in learning low-dimensional latent features of high-dimensional data across various applications. Assuming that data are sampled near a low-dimensional manifold, we employ chart autoencoders, which encode data into low-dimensional latent features on a collection of charts, preserving the topology and geometry of the data manifold. Our paper establishes statistical guarantees on the generalization error of chart autoencoders, and we demonstrate their denoising capabilities by considering $n$ noisy training samples, along with their noise-free counterparts, on a $d$-dimensional manifold. By training autoencoders, we show that chart autoencoders can effectively denoise the input data with normal noise. We prove that, under proper network architectures, chart autoencoders achieve a squared generalization error in the order of $\displaystyle n^{-\frac{2}{d+2}}\log^4 n$, which depends on the intrinsic dimension of the manifold and only weakly depends on the ambient dimension and noise level. We further extend our theory on data with noise containing both normal and tangential components, where chart autoencoders still exhibit a denoising effect for the normal component. As a special case, our theory also applies to classical autoencoders, as long as the data manifold has a global parametrization. Our results provide a solid theoretical foundation for the effectiveness of autoencoders, which is further validated through several numerical experiments.
The pre-training and fine-tuning paradigm has contributed to a number of breakthroughs in Natural Language Processing (NLP). Instead of directly training on a downstream task, language models are first pre-trained on large datasets with cross-domain knowledge (e.g., Pile, MassiveText, etc.) and then fine-tuned on task-specific data (e.g., natural language generation, text summarization, etc.). Scaling the model and dataset size has helped improve the performance of LLMs, but unfortunately, this also leads to highly prohibitive computational costs. Pre-training LLMs often require orders of magnitude more FLOPs than fine-tuning and the model capacity often remains the same between the two phases. To achieve training efficiency w.r.t training FLOPs, we propose to decouple the model capacity between the two phases and introduce Sparse Pre-training and Dense Fine-tuning (SPDF). In this work, we show the benefits of using unstructured weight sparsity to train only a subset of weights during pre-training (Sparse Pre-training) and then recover the representational capacity by allowing the zeroed weights to learn (Dense Fine-tuning). We demonstrate that we can induce up to 75% sparsity into a 1.3B parameter GPT-3 XL model resulting in a 2.5x reduction in pre-training FLOPs, without a significant loss in accuracy on the downstream tasks relative to the dense baseline. By rigorously evaluating multiple downstream tasks, we also establish a relationship between sparsity, task complexity, and dataset size. Our work presents a promising direction to train large GPT models at a fraction of the training FLOPs using weight sparsity while retaining the benefits of pre-trained textual representations for downstream tasks.
Saliency detection methods are central to several real-world applications such as robot navigation and satellite imagery. However, the performance of existing methods deteriorate under low-light conditions because training datasets mostly comprise of well-lit images. One possible solution is to collect a new dataset for low-light conditions. This involves pixel-level annotations, which is not only tedious and time-consuming but also infeasible if a huge training corpus is required. We propose a technique that performs classical band-pass filtering in the Fourier space to transform well-lit images to low-light images and use them as a proxy for real low-light images. Unlike popular deep learning approaches which require learning thousands of parameters and enormous amounts of training data, the proposed transformation is fast and simple and easy to extend to other tasks such as low-light depth estimation. Our experiments show that the state-of-the-art saliency detection and depth estimation networks trained on our proxy low-light images perform significantly better on real low-light images than networks trained using existing strategies.
An instrumental variable (IV) is a device that encourages units in a study to be exposed to a treatment. Under a set of key assumptions, a valid instrument allows for consistent estimation of treatment effects for compliers (those who are only exposed to treatment when encouraged to do so) even in the presence of unobserved confounders. Unfortunately, popular IV estimators can be unstable in studies with a small fraction of compliers. Here, we explore post-stratifying the data using variables that predict complier status (and, potentially, the outcome) to yield better estimation and inferential properties. We outline an estimator that is a weighted average of IV estimates within each stratum, weighing the stratum estimates by their estimated proportion of compliers. We then explore the benefits of post-stratification in terms of bias reduction, variance reduction, and improved standard error estimates, providing derivations that identify the direction of bias as a function of the relative means of the compliers and non-compliers. We also provide a finite-sample asymptotic formula for the variance of the post-stratified estimators. We demonstrate the relative performances of different IV approaches in simulations studies and discuss the advantages of our design-based post-stratification approach over incorporating compliance-predictive covariates into two-stage least squares regressions. In the end, we show covariates predictive of outcome can increase precision, but only if one is willing to make a bias-variance trade-off by down-weighting or dropping those strata with few compliers. Our methods are further exemplified in an application.
Recent years have witnessed a remarkable success of large deep learning models. However, training these models is challenging due to high computational costs, painfully slow convergence, and overfitting issues. In this paper, we present Deep Incubation, a novel approach that enables the efficient and effective training of large models by dividing them into smaller sub-modules that can be trained separately and assembled seamlessly. A key challenge for implementing this idea is to ensure the compatibility of the independently trained sub-modules. To address this issue, we first introduce a global, shared meta model, which is leveraged to implicitly link all the modules together, and can be designed as an extremely small network with negligible computational overhead. Then we propose a module incubation algorithm, which trains each sub-module to replace the corresponding component of the meta model and accomplish a given learning task. Despite the simplicity, our approach effectively encourages each sub-module to be aware of its role in the target large model, such that the finally-learned sub-modules can collaborate with each other smoothly after being assembled. Empirically, our method outperforms end-to-end (E2E) training in terms of both final accuracy and training efficiency. For example, on top of ViT-Huge, it improves the accuracy by 2.7% on ImageNet or achieves similar performance with 4x less training time. Notably, the gains are significant for downstream tasks as well (e.g., object detection and image segmentation on COCO and ADE20K). Code is available at //github.com/LeapLabTHU/Deep-Incubation.
Multimodal hate detection, which aims to identify harmful content online such as memes, is crucial for building a wholesome internet environment. Previous work has made enlightening exploration in detecting explicit hate remarks. However, most of their approaches neglect the analysis of implicit harm, which is particularly challenging as explicit text markers and demographic visual cues are often twisted or missing. The leveraged cross-modal attention mechanisms also suffer from the distributional modality gap and lack logical interpretability. To address these semantic gaps issues, we propose TOT: a topology-aware optimal transport framework to decipher the implicit harm in memes scenario, which formulates the cross-modal aligning problem as solutions for optimal transportation plans. Specifically, we leverage an optimal transport kernel method to capture complementary information from multiple modalities. The kernel embedding provides a non-linear transformation ability to reproduce a kernel Hilbert space (RKHS), which reflects significance for eliminating the distributional modality gap. Moreover, we perceive the topology information based on aligned representations to conduct bipartite graph path reasoning. The newly achieved state-of-the-art performance on two publicly available benchmark datasets, together with further visual analysis, demonstrate the superiority of TOT in capturing implicit cross-modal alignment.
Extreme multi-label text classification (XMC) aims to tag each input text with the most relevant labels from an extremely large label set, such as those that arise in product categorization and e-commerce recommendation. Recently, pretrained language representation models such as BERT achieve remarkable state-of-the-art performance across a wide range of NLP tasks including sentence classification among small label sets (typically fewer than thousands). Indeed, there are several challenges in applying BERT to the XMC problem. The main challenges are: (i) the difficulty of capturing dependencies and correlations among labels, whose features may come from heterogeneous sources, and (ii) the tractability to scale to the extreme label setting as the model size can be very large and scale linearly with the size of the output space. To overcome these challenges, we propose X-BERT, the first feasible attempt to finetune BERT models for a scalable solution to the XMC problem. Specifically, X-BERT leverages both the label and document text to build label representations, which induces semantic label clusters in order to better model label dependencies. At the heart of X-BERT is finetuning BERT models to capture the contextual relations between input text and the induced label clusters. Finally, an ensemble of the different BERT models trained on heterogeneous label clusters leads to our best final model. Empirically, on a Wiki dataset with around 0.5 million labels, X-BERT achieves new state-of-the-art results where the precision@1 reaches 67:80%, a substantial improvement over 32.58%/60.91% of deep learning baseline fastText and competing XMC approach Parabel, respectively. This amounts to a 11.31% relative improvement over Parabel, which is indeed significant since the recent approach SLICE only has 5.53% relative improvement.
Recently, ensemble has been applied to deep metric learning to yield state-of-the-art results. Deep metric learning aims to learn deep neural networks for feature embeddings, distances of which satisfy given constraint. In deep metric learning, ensemble takes average of distances learned by multiple learners. As one important aspect of ensemble, the learners should be diverse in their feature embeddings. To this end, we propose an attention-based ensemble, which uses multiple attention masks, so that each learner can attend to different parts of the object. We also propose a divergence loss, which encourages diversity among the learners. The proposed method is applied to the standard benchmarks of deep metric learning and experimental results show that it outperforms the state-of-the-art methods by a significant margin on image retrieval tasks.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191