Temporally consistent depth estimation is crucial for real-time applications such as augmented reality. While stereo depth estimation has received substantial attention that led to improvements on a frame-by-frame basis, there is relatively little work focused on maintaining temporal consistency across frames. Indeed, based on our analysis, current stereo depth estimation techniques still suffer from poor temporal consistency. Stabilizing depth temporally in dynamic scenes is challenging due to concurrent object and camera motion. In an online setting, this process is further aggravated because only past frames are available. In this paper, we present a technique to produce temporally consistent depth estimates in dynamic scenes in an online setting. Our network augments current per-frame stereo networks with novel motion and fusion networks. The motion network accounts for both object and camera motion by predicting a per-pixel SE3 transformation. The fusion network improves consistency in prediction by aggregating the current and previous predictions with regressed weights. We conduct extensive experiments across varied datasets (synthetic, outdoor, indoor and medical). In both zero-shot generalization and domain fine-tuning, we demonstrate that our proposed approach outperforms competing methods in terms of temporal stability and per-frame accuracy, both quantitatively and qualitatively. Our code will be available online.
Vision transformers (ViTs) have attracted much attention for their superior performance on computer vision tasks. To address their limitations of single-scale low-resolution representations, prior work adapts ViTs to high-resolution dense prediction tasks with hierarchical architectures to generate pyramid features. However, multi-scale representation learning is still under-explored on ViTs, given their classification-like sequential topology. To enhance ViTs with more capability to learn semantically-rich and spatially-precise multi-scale representations, in this work, we present an efficient integration of high-resolution multi-branch architectures with vision transformers, dubbed HRViT, pushing the Pareto front of dense prediction tasks to a new level. We explore heterogeneous branch design, reduce the redundancy in linear layers, and augment the model nonlinearity to balance the model performance and hardware efficiency. The proposed HRViT achieves 50.20% mIoU on ADE20K and 83.16% mIoU on Cityscapes for semantic segmentation tasks, surpassing state-of-the-art MiT and CSWin with an average of +1.78 mIoU improvement, 28% parameter reduction, and 21% FLOPs reduction, demonstrating the potential of HRViT as strong vision backbones.
From wearables to powerful smart devices, modern automatic speech recognition (ASR) models run on a variety of edge devices with different computational budgets. To navigate the Pareto front of model accuracy vs model size, researchers are trapped in a dilemma of optimizing model accuracy by training and fine-tuning models for each individual edge device while keeping the training GPU-hours tractable. In this paper, we propose Omni-sparsity DNN, where a single neural network can be pruned to generate optimized model for a large range of model sizes. We develop training strategies for Omni-sparsity DNN that allows it to find models along the Pareto front of word-error-rate (WER) vs model size while keeping the training GPU-hours to no more than that of training one singular model. We demonstrate the Omni-sparsity DNN with streaming E2E ASR models. Our results show great saving on training time and resources with similar or better accuracy on LibriSpeech compared to individually pruned sparse models: 2%-6.6% better WER on Test-other.
This paper improves the streaming transformer transducer for speech recognition by using non-causal convolution. Many works apply the causal convolution to improve streaming transformer ignoring the lookahead context. We propose to use non-causal convolution to process the center block and lookahead context separately. This method leverages the lookahead context in convolution and maintains similar training and decoding efficiency. Given the similar latency, using the non-causal convolution with lookahead context gives better accuracy than causal convolution, especially for open-domain dictation scenarios. Besides, this paper applies talking-head attention and a novel history context compression scheme to further improve the performance. The talking-head attention improves the multi-head self-attention by transferring information among different heads. The history context compression method introduces more extended history context compactly. On our in-house data, the proposed methods improve a small Emformer baseline with lookahead context by relative WERR 5.1\%, 14.5\%, 8.4\% on open-domain dictation, assistant general scenarios, and assistant calling scenarios, respectively.
Automatic speech recognition (ASR) has become increasingly ubiquitous on modern edge devices. Past work developed streaming End-to-End (E2E) all-neural speech recognizers that can run compactly on edge devices. However, E2E ASR models are prone to overfitting and have difficulties in generalizing to unseen testing data. Various techniques have been proposed to regularize the training of ASR models, including layer normalization, dropout, spectrum data augmentation and speed distortions in the inputs. In this work, we present a simple yet effective noisy training strategy to further improve the E2E ASR model training. By introducing random noise to the parameter space during training, our method can produce smoother models at convergence that generalize better. We apply noisy training to improve both dense and sparse state-of-the-art Emformer models and observe consistent WER reduction. Specifically, when training Emformers with 90% sparsity, we achieve 12% and 14% WER improvements on the LibriSpeech Test-other and Test-clean data set, respectively.
Introducing the transformer structure into computer vision tasks holds the promise of yielding a better speed-accuracy trade-off than traditional convolution networks. However, directly training vanilla transformers on vision tasks has been shown to yield unstable and sub-optimal results. As a result, recent works propose to modify transformer structures by incorporating convolutional layers to improve the performance on vision tasks. This work investigates how to stabilize the training of vision transformers \emph{without} special structure modification. We observe that the instability of transformer training on vision tasks can be attributed to the over-smoothing problem, that the self-attention layers tend to map the different patches from the input image into a similar latent representation, hence yielding the loss of information and degeneration of performance, especially when the number of layers is large. We then propose a number of techniques to alleviate this problem, including introducing additional loss functions to encourage diversity, prevent loss of information, and discriminate different patches by additional patch classification loss for Cutmix. We show that our proposed techniques stabilize the training and allow us to train wider and deeper vision transformers, achieving 85.0\% top-1 accuracy on ImageNet validation set without introducing extra teachers or additional convolution layers. Our code will be made publicly available at https://github.com/ChengyueGongR/PatchVisionTransformer .
Weight-sharing neural architecture search (NAS) is an effective technique for automating efficient neural architecture design. Weight-sharing NAS builds a supernet that assembles all the architectures as its sub-networks and jointly trains the supernet with the sub-networks. The success of weight-sharing NAS heavily relies on distilling the knowledge of the supernet to the sub-networks. However, we find that the widely used distillation divergence, i.e., KL divergence, may lead to student sub-networks that over-estimate or under-estimate the uncertainty of the teacher supernet, leading to inferior performance of the sub-networks. In this work, we propose to improve the supernet training with a more generalized alpha-divergence. By adaptively selecting the alpha-divergence, we simultaneously prevent the over-estimation or under-estimation of the uncertainty of the teacher model. We apply the proposed alpha-divergence based supernet training to both slimmable neural networks and weight-sharing NAS, and demonstrate significant improvements. Specifically, our discovered model family, AlphaNet, outperforms prior-art models on a wide range of FLOPs regimes, including BigNAS, Once-for-All networks, FBNetV3, and AttentiveNAS. We achieve ImageNet top-1 accuracy of 80.0% with only 444 MFLOPs.
Semi-supervised learning (SSL) is a key approach toward more data-efficient machine learning by jointly leverage both labeled and unlabeled data. We propose AlphaMatch, an efficient SSL method that leverages data augmentations, by efficiently enforcing the label consistency between the data points and the augmented data derived from them. Our key technical contribution lies on: 1) using alpha-divergence to prioritize the regularization on data with high confidence, achieving a similar effect as FixMatch but in a more flexible fashion, and 2) proposing an optimization-based, EM-like algorithm to enforce the consistency, which enjoys better convergence than iterative regularization procedures used in recent SSL methods such as FixMatch, UDA, and MixMatch. AlphaMatch is simple and easy to implement, and consistently outperforms prior arts on standard benchmarks, e.g. CIFAR-10, SVHN, CIFAR-100, STL-10. Specifically, we achieve 91.3% test accuracy on CIFAR-10 with just 4 labelled data per class, substantially improving over the previously best 88.7% accuracy achieved by FixMatch.
Data augmentation (DA) is an essential technique for training state-of-the-art deep learning systems. In this paper, we empirically show data augmentation might introduce noisy augmented examples and consequently hurt the performance on unaugmented data during inference. To alleviate this issue, we propose a simple yet highly effective approach, dubbed \emph{KeepAugment}, to increase augmented images fidelity. The idea is first to use the saliency map to detect important regions on the original images and then preserve these informative regions during augmentation. This information-preserving strategy allows us to generate more faithful training examples. Empirically, we demonstrate our method significantly improves on a number of prior art data augmentation schemes, e.g. AutoAugment, Cutout, random erasing, achieving promising results on image classification, semi-supervised image classification, multi-view multi-camera tracking and object detection.
Neural architecture search (NAS) has shown great promise designing state-of-the-art (SOTA) models that are both accurate and fast. Recently, two-stage NAS, e.g. BigNAS, decouples the model training and searching process and achieves good search efficiency. Two-stage NAS requires sampling from the search space during training, which directly impacts the accuracy of the final searched models. While uniform sampling has been widely used for simplicity, it is agnostic of the model performance Pareto front, which are the main focus in the search process, and thus, misses opportunities to further improve the model accuracy. In this work, we propose AttentiveNAS that focuses on sampling the networks to improve the performance Pareto. We also propose algorithms to efficiently and effectively identify the networks on the Pareto during training. Without extra re-training or post-processing, we can simultaneously obtain a large number of networks across a wide range of FLOPs. Our discovered model family, AttentiveNAS models, achieves top-1 accuracy from 77.3% to 80.7% on ImageNet, and outperforms SOTA models, including BigNAS, Once-for-All networks and FBNetV3. We also achieve ImageNet accuracy of 80.1% with only 491 MFLOPs.