Video Denoising is one of the fundamental tasks of any videoprocessing pipeline. It is different from image denoising due to the tem-poral aspects of video frames, and any image denoising approach appliedto videos will result in flickering. The proposed method makes use oftemporal as well as spatial dimensions of video frames as part of a two-stage pipeline. Each stage in the architecture named as Spatio-TemporalNetwork uses a channel-wise attention mechanism to forward the encodersignal to the decoder side. The Attention Block used in this paper usessoft attention to ranks the filters for better training.
We consider the design of two-pass voice trigger detection systems. We focus on the networks in the second pass that are used to re-score candidate segments obtained from the first-pass. Our baseline is an acoustic model(AM), with BiLSTM layers, trained by minimizing the CTC loss. We replace the BiLSTM layers with self-attention layers. Results on internal evaluation sets show that self-attention networks yield better accuracy while requiring fewer parameters. We add an auto-regressive decoder network on top of the self-attention layers and jointly minimize the CTC loss on the encoder and the cross-entropy loss on the decoder. This design yields further improvements over the baseline. We retrain all the models above in a multi-task learning(MTL) setting, where one branch of a shared network is trained as an AM, while the second branch classifies the whole sequence to be true-trigger or not. Results demonstrate that networks with self-attention layers yield $\sim$60% relative reduction in false reject rates for a given false-alarm rate, while requiring 10% fewer parameters. When trained in the MTL setup, self-attention networks yield further accuracy improvements. On-device measurements show that we observe 70% relative reduction in inference time. Additionally, the proposed network architectures are $\sim$5X faster to train.
In this study, a deep-learning-based multi-stage network architecture called Multi-Stage Depth Prediction Network (MSDPN) is proposed to predict a dense depth map using a 2D LiDAR and a monocular camera. Our proposed network consists of a multi-stage encoder-decoder architecture and Cross Stage Feature Aggregation (CSFA). The proposed multi-stage encoder-decoder architecture alleviates the partial observation problem caused by the characteristics of a 2D LiDAR, and CSFA prevents the multi-stage network from diluting the features and allows the network to learn the inter-spatial relationship between features better. Previous works use sub-sampled data from the ground truth as an input rather than actual 2D LiDAR data. In contrast, our approach trains the model and conducts experiments with a physically-collected 2D LiDAR dataset. To this end, we acquired our own dataset called KAIST RGBD-scan dataset and validated the effectiveness and the robustness of MSDPN under realistic conditions. As verified experimentally, our network yields promising performance against state-of-the-art methods. Additionally, we analyzed the performance of different input methods and confirmed that the reference depth map is robust in untrained scenarios.
Deep neural nets typically perform end-to-end backpropagation to learn the weights, a procedure that creates synchronization constraints in the weight update step across layers and is not biologically plausible. Recent advances in unsupervised contrastive representation learning point to the question of whether a learning algorithm can also be made local, that is, the updates of lower layers do not directly depend on the computation of upper layers. While Greedy InfoMax separately learns each block with a local objective, we found that it consistently hurts readout accuracy in state-of-the-art unsupervised contrastive learning algorithms, possibly due to the greedy objective as well as gradient isolation. In this work, we discover that by overlapping local blocks stacking on top of each other, we effectively increase the decoder depth and allow upper blocks to implicitly send feedbacks to lower blocks. This simple design closes the performance gap between local learning and end-to-end contrastive learning algorithms for the first time. Aside from standard ImageNet experiments, we also show results on complex downstream tasks such as object detection and instance segmentation directly using readout features.
We propose a novel algorithm, named Open-Edit, which is the first attempt on open-domain image manipulation with open-vocabulary instructions. It is a challenging task considering the large variation of image domains and the lack of training supervision. Our approach takes advantage of the unified visual-semantic embedding space pretrained on a general image-caption dataset, and manipulates the embedded visual features by applying text-guided vector arithmetic on the image feature maps. A structure-preserving image decoder then generates the manipulated images from the manipulated feature maps. We further propose an on-the-fly sample-specific optimization approach with cycle-consistency constraints to regularize the manipulated images and force them to preserve details of the source images. Our approach shows promising results in manipulating open-vocabulary color, texture, and high-level attributes for various scenarios of open-domain images.
In this paper, we study a federated learning system at the wireless edge that uses over-the-air computation (AirComp). In such a system, users transmit their messages over a multi-access channel concurrently to achieve fast model aggregation. Recently, an AirComp scheme based on digital modulation has been proposed featuring one-bit gradient quantization and truncated channel inversion at users and a majority-voting based decoder at the fusion center (FC). We propose an improved digital AirComp scheme to relax its requirements on the transmitters, where users perform phase correction and transmit with full power. To characterize the decoding failure probability at the FC, we introduce the normalized detection signal-to-noise ratio (SNR), which can be interpreted as the effective participation rate of users. To mitigate wireless fading, we further propose a cluster-based system and design the relay selection scheme based on the normalized detection SNR. By local data fusion within each cluster and relay selection, our scheme can fully exploit spatial diversity to increase the effective number of voting users and accelerate model convergence.
Learning useful representations of complex data has been the subject of extensive research for many years. With the diffusion of Deep Neural Networks, Variational Autoencoders have gained lots of attention since they provide an explicit model of the data distribution based on an encoder/decoder architecture which is able to both generate images and encode them in a low-dimensional subspace. However, the latent space is not easily interpretable and the generation capabilities show some limitations since images typically look blurry and lack details. In this paper, we propose the Introspective Variational Classifier (IntroVAC), a model that learns interpretable latent subspaces by exploiting information from an additional label and provides improved image quality thanks to an adversarial training strategy.We show that IntroVAC is able to learn meaningful directions in the latent space enabling fine-grained manipulation of image attributes. We validate our approach on the CelebA dataset.
Handling digital images is almost always accompanied by a lossy compression in order to facilitate efficient transmission and storage. This introduces an unavoidable tension between the allocated bit-budget (rate) and the faithfulness of the resulting image to the original one (distortion). An additional complicating consideration is the effect of the compression on recognition performance by given classifiers (accuracy). This work aims to explore this rate-distortion-accuracy tradeoff. As a case study, we focus on the design of the quantization tables in the JPEG compression standard. We offer a novel optimal tuning of these tables via continuous optimization, leveraging a differential implementation of both the JPEG encoder-decoder and an entropy estimator. This enables us to offer a unified framework that considers the interplay between rate, distortion and classification accuracy. In all these fronts, we report a substantial boost in performance by a simple and easily implemented modification of these tables.
This paper proposes an encoder-decoder architecture grounded on Generative Adversarial Networks (GANs), that generates a sequence of activities and their timestamps in an end-to-end way. GANs work well with differentiable data such as images. However, a suffix is a sequence of categorical items. To this end, we use the Gumbel-Softmax distribution to get a differentiable continuous approximation. The training works by putting one neural network against the other in a two-player game (hence the "adversarial" nature), which leads to generating suffixes close to the ground truth. From the experimental evaluation it emerges that the approach is superior to the baselines in terms of the accuracy of the predicted suffixes and corresponding remaining times, despite using a naive feature encoding and only engineering features based on control flow and events completion time.
Machine translation has many applications such as news translation, email translation, official letter translation etc. Commercial translators, e.g. Google Translation lags in regional vocabulary and are unable to learn the bilingual text in the source and target languages within the input. In this paper, a regional vocabulary-based application-oriented Neural Machine Translation (NMT) model is proposed over the data set of emails used at the University for communication over a period of three years. A state-of-the-art Sequence-to-Sequence Neural Network for ML -> EN and EN -> ML translations is compared with Google Translate using Gated Recurrent Unit Recurrent Neural Network machine translation model with attention decoder. The low BLEU score of Google Translation in comparison to our model indicates that the application based regional models are better. The low BLEU score of EN -> ML of our model and Google Translation indicates that the Malay Language has complex language features corresponding to English.
In order to perform network analysis tasks, representations that capture the most relevant information in the graph structure are needed. However, existing methods do not learn representations that can be interpreted in a straightforward way and that are robust to perturbations to the graph structure. In this work, we address these two limitations by proposing node2coords, a representation learning algorithm for graphs, which learns simultaneously a low-dimensional space and coordinates for the nodes in that space. The patterns that span the low dimensional space reveal the graph's most important structural information. The coordinates of the nodes reveal the proximity of their local structure to the graph structural patterns. In order to measure this proximity by taking into account the underlying graph, we propose to use Wasserstein distances. We introduce an autoencoder that employs a linear layer in the encoder and a novel Wasserstein barycentric layer at the decoder. Node connectivity descriptors, that capture the local structure of the nodes, are passed through the encoder to learn the small set of graph structural patterns. In the decoder, the node connectivity descriptors are reconstructed as Wasserstein barycenters of the graph structural patterns. The optimal weights for the barycenter representation of a node's connectivity descriptor correspond to the coordinates of that node in the low-dimensional space. Experimental results demonstrate that the representations learned with node2coords are interpretable, lead to node embeddings that are stable to perturbations of the graph structure and achieve competitive or superior results compared to state-of-the-art methods in node classification.
In this paper we present an end-to-end meta-learned system for image compression. Traditional machine learning based approaches to image compression train one or more neural network for generalization performance. However, at inference time, the encoder or the latent tensor output by the encoder can be optimized for each test image. This optimization can be regarded as a form of adaptation or benevolent overfitting to the input content. In order to reduce the gap between training and inference conditions, we propose a new training paradigm for learned image compression, which is based on meta-learning. In a first phase, the neural networks are trained normally. In a second phase, the Model-Agnostic Meta-learning approach is adapted to the specific case of image compression, where the inner-loop performs latent tensor overfitting, and the outer loop updates both encoder and decoder neural networks based on the overfitting performance. Furthermore, after meta-learning, we propose to overfit and cluster the bias terms of the decoder on training image patches, so that at inference time the optimal content-specific bias terms can be selected at encoder-side. Finally, we propose a new probability model for lossless compression, which combines concepts from both multi-scale and super-resolution probability model approaches. We show the benefits of all our proposed ideas via carefully designed experiments.
In this paper, we propose a cascaded non-local neural network for point cloud segmentation. The proposed network aims to build the long-range dependencies of point clouds for the accurate segmentation. Specifically, we develop a novel cascaded non-local module, which consists of the neighborhood-level, superpoint-level and global-level non-local blocks. First, in the neighborhood-level block, we extract the local features of the centroid points of point clouds by assigning different weights to the neighboring points. The extracted local features of the centroid points are then used to encode the superpoint-level block with the non-local operation. Finally, the global-level block aggregates the non-local features of the superpoints for semantic segmentation in an encoder-decoder framework. Benefiting from the cascaded structure, geometric structure information of different neighborhoods with the same label can be propagated. In addition, the cascaded structure can largely reduce the computational cost of the original non-local operation on point clouds. Experiments on different indoor and outdoor datasets show that our method achieves state-of-the-art performance and effectively reduces the time consumption and memory occupation.
Supervised approaches for Neural Abstractive Summarization require large annotated corpora that are costly to build. We present a French meeting summarization task where reports are predicted based on the automatic transcription of the meeting audio recordings. In order to build a corpus for this task, it is necessary to obtain the (automatic or manual) transcription of each meeting, and then to segment and align it with the corresponding manual report to produce training examples suitable for training. On the other hand, we have access to a very large amount of unaligned data, in particular reports without corresponding transcription. Reports are professionally written and well formatted making pre-processing straightforward. In this context, we study how to take advantage of this massive amount of unaligned data using two approaches (i) self-supervised pre-training using a target-side denoising encoder-decoder model; (ii) back-summarization i.e. reversing the summarization process by learning to predict the transcription given the report, in order to align single reports with generated transcription, and use this synthetic dataset for further training. We report large improvements compared to the previous baseline (trained on aligned data only) for both approaches on two evaluation sets. Moreover, combining the two gives even better results, outperforming the baseline by a large margin of +6 ROUGE-1 and ROUGE-L and +5 ROUGE-2 on two evaluation sets
We unify empirical Bayes and variational Bayes for approximating unnormalized densities. This framework, named unnormalized variational Bayes (UVB), is based on formulating a latent variable model for the random variable $Y=X+N(0,\sigma^2 I_d)$ and using the evidence lower bound (ELBO), computed by a variational autoencoder, as a parametrization of the energy function of $Y$ which is then used to estimate $X$ with the empirical Bayes least-squares estimator. In this intriguing setup, the $\textit{gradient}$ of the ELBO with respect to noisy inputs plays the central role in learning the energy function. Empirically, we demonstrate that UVB has a higher capacity to approximate energy functions than the parametrization with MLPs as done in neural empirical Bayes (DEEN). We especially showcase $\sigma=1$, where the differences between UVB and DEEN become visible and qualitative in the denoising experiments. For this high level of noise, the distribution of $Y$ is very smoothed and we demonstrate that one can traverse in a single run $-$ without a restart $-$ all MNIST classes in a variety of styles via walk-jump sampling with a fast-mixing Langevin MCMC sampler. We finish by probing the encoder/decoder of the trained models and confirm UVB $\neq$ VAE.
It has recently been demonstrated that spatial resolution adaptation can be integrated within video compression to improve overall coding performance by spatially down-sampling before encoding and super-resolving at the decoder. Significant improvements have been reported when convolutional neural networks (CNNs) were used to perform the resolution up-sampling. However, this approach suffers from high complexity at the decoder due to the employment of CNN-based super-resolution. In this paper, a novel framework is proposed which supports the flexible allocation of complexity between the encoder and decoder. This approach employs a CNN model for video down-sampling at the encoder and uses a Lanczos3 filter to reconstruct full resolution at the decoder. The proposed method was integrated into the HEVC HM 16.20 software and evaluated on JVET UHD test sequences using the All Intra configuration. The experimental results demonstrate the potential of the proposed approach, with significant bitrate savings (more than 10%) over the original HEVC HM, coupled with reduced computational complexity at both encoder (29%) and decoder (10%).
Pedestrian trajectory prediction is an essential task in robotic applications such as autonomous driving and robot navigation. State-of-the-art trajectory predictors use a conditional variational autoencoder (CVAE) with recurrent neural networks (RNNs) to encode observed trajectories and decode multi-modal future trajectories. This process can suffer from accumulated errors over long prediction horizons (>=2 seconds). This paper presents BiTraP, a goal-conditioned bi-directional multi-modal trajectory prediction method based on the CVAE. BiTraP estimates the goal (end-point) of trajectories and introduces a novel bi-directional decoder to improve longer-term trajectory prediction accuracy. Extensive experiments show that BiTraP generalizes to both first-person view (FPV) and bird's-eye view (BEV) scenarios and outperforms state-of-the-art results by ~10-50%. We also show that different choices of non-parametric versus parametric target models in the CVAE directly influence the predicted multi-modal trajectory distributions. These results provide guidance on trajectory predictor design for robotic applications such as collision avoidance and navigation systems.
In the present paper, we propose a decoder-free extension of Dreamer, a leading model-based reinforcement learning (MBRL) method from pixels. Dreamer is a sample- and cost-efficient solution to robot learning, as it is used to train latent state-space models based on a variational autoencoder and to conduct policy optimization by latent trajectory imagination. However, this autoencoding based approach often causes object vanishing, in which the autoencoder fails to perceives key objects for solving control tasks, and thus significantly limiting Dreamer's potential. This work aims to relieve this Dreamer's bottleneck and enhance its performance by means of removing the decoder. For this purpose, we firstly derive a likelihood-free and InfoMax objective of contrastive learning from the evidence lower bound of Dreamer. Secondly, we incorporate two components, (i) independent linear dynamics and (ii) the random crop data augmentation, to the learning scheme so as to improve the training performance. In comparison to Dreamer and other recent model-free reinforcement learning methods, our newly devised Dreamer with InfoMax and without generative decoder (Dreaming) achieves the best scores on 5 difficult simulated robotics tasks, in which Dreamer suffers from object vanishing.
Deep learning is a hot research topic in the field of machine learning methods and applications. Generative Adversarial Networks (GANs) and Variational Auto-Encoders (VAEs) provide impressive image generations from Gaussian white noise, but both of them are difficult to train since they need to train the generator (or encoder) and the discriminator (or decoder) simultaneously, which is easy to cause unstable training. In order to solve or alleviate the synchronous training difficult problems of GANs and VAEs, recently, researchers propose Generative Scattering Networks (GSNs), which use wavelet scattering networks (ScatNets) as the encoder to obtain the features (or ScatNet embeddings) and convolutional neural networks (CNNs) as the decoder to generate the image. The advantage of GSNs is the parameters of ScatNets are not needed to learn, and the disadvantage of GSNs is that the expression ability of ScatNets is slightly weaker than CNNs and the dimensional reduction method of Principal Component Analysis (PCA) is easy to lead overfitting in the training of GSNs, and therefore affect the generated quality in the testing process. In order to further improve the quality of generated images while keep the advantages of GSNs, this paper proposes Generative Fractional Scattering Networks (GFRSNs), which use more expressive fractional wavelet scattering networks (FrScatNets) instead of ScatNets as the encoder to obtain the features (or FrScatNet embeddings) and use the similar CNNs of GSNs as the decoder to generate the image. Additionally, this paper develops a new dimensional reduction method named Feature-Map Fusion (FMF) instead of PCA for better keeping the information of FrScatNets and the effect of image fusion on the quality of image generation is also discussed.
Unsupervised active learning has attracted increasing attention in recent years, where its goal is to select representative samples in an unsupervised setting for human annotating. Most existing works are based on shallow linear models by assuming that each sample can be well approximated by the span (i.e., the set of all linear combinations) of certain selected samples, and then take these selected samples as representative ones to label. However, in practice, the data do not necessarily conform to linear models, and how to model nonlinearity of data often becomes the key point to success. In this paper, we present a novel Deep neural network framework for Unsupervised Active Learning, called DUAL. DUAL can explicitly learn a nonlinear embedding to map each input into a latent space through an encoder-decoder architecture, and introduce a selection block to select representative samples in the the learnt latent space. In the selection block, DUAL considers to simultaneously preserve the whole input patterns as well as the cluster structure of data. Extensive experiments are performed on six publicly available datasets, and experimental results clearly demonstrate the efficacy of our method, compared with state-of-the-arts.