Jungwon Lee

Junhyuk So, Jungwon Lee, Daehyun Ahn et al.

The diffusion model has gained popularity in vision applications due to its remarkable generative performance and versatility. However, high storage and computation demands, resulting from the model size and iterative generation, hinder its use on mobile devices. Existing quantization techniques struggle to maintain performance even in 8-bit precision due to the diffusion model's unique property of temporal variation in activation. We introduce a novel quantization method that dynamically adjusts the quantization interval based on time step information, significantly improving output quality. Unlike conventional dynamic quantization techniques, our approach has no computational overhead during inference and is compatible with both post-training quantization (PTQ) and quantization-aware training (QAT). Our extensive experiments demonstrate substantial improvements in output quality with the quantized diffusion model across various datasets.

Yoojin Choi, Mostafa El-Khamy, Jungwon Lee

We propose a novel method for training a conditional generative adversarial network (CGAN) without the use of training data, called zero-shot learning of a CGAN (ZS-CGAN). Zero-shot learning of a conditional generator only needs a pre-trained discriminative (classification) model and does not need any training data. In particular, the conditional generator is trained to produce labeled synthetic samples whose characteristics mimic the original training data by using the statistics stored in the batch normalization layers of the pre-trained model. We show the usefulness of ZS-CGAN in data-free quantization of deep neural networks. We achieved the state-of-the-art data-free network quantization of the ResNet and MobileNet classification models trained on the ImageNet dataset. Data-free quantization using ZS-CGAN showed a minimal loss in accuracy compared to that obtained by conventional data-dependent quantization.

Junhyuk So, Jungwon Lee, Eunhyeok Park

The substantial computational costs of diffusion models, especially due to the repeated denoising steps necessary for high-quality image generation, present a major obstacle to their widespread adoption. While several studies have attempted to address this issue by reducing the number of score function evaluations (NFE) using advanced ODE solvers without fine-tuning, the decreased number of denoising iterations misses the opportunity to update fine details, resulting in noticeable quality degradation. In our work, we introduce an advanced acceleration technique that leverages the temporal redundancy inherent in diffusion models. Reusing feature maps with high temporal similarity opens up a new opportunity to save computation resources without compromising output quality. To realize the practical benefits of this intuition, we conduct an extensive analysis and propose a novel method, FRDiff. FRDiff is designed to harness the advantages of both reduced NFE and feature reuse, achieving a Pareto frontier that balances fidelity and latency trade-offs in various generative tasks.

Yoojin Choi, Mostafa El-Khamy, Jungwon Lee

This paper proposes two novel knowledge transfer techniques for class-incremental learning (CIL). First, we propose data-free generative replay (DF-GR) to mitigate catastrophic forgetting in CIL by using synthetic samples from a generative model. In the conventional generative replay, the generative model is pre-trained for old data and shared in extra memory for later incremental learning. In our proposed DF-GR, we train a generative model from scratch without using any training data, based on the pre-trained classification model from the past, so we curtail the cost of sharing pre-trained generative models. Second, we introduce dual-teacher information distillation (DT-ID) for knowledge distillation from two teachers to one student. In CIL, we use DT-ID to learn new classes incrementally based on the pre-trained model for old classes and another model (pre-)trained on the new data for new classes. We implemented the proposed schemes on top of one of the state-of-the-art CIL methods and showed the performance improvement on CIFAR-100 and ImageNet datasets.

Weituo Hao, Mostafa El-Khamy, Jungwon Lee et al.

Federated learning has emerged as an important distributed learning paradigm, where a server aggregates a global model from many client-trained models while having no access to the client data. Although it is recognized that statistical heterogeneity of the client local data yields slower global model convergence, it is less commonly recognized that it also yields a biased federated global model with a high variance of accuracy across clients. In this work, we aim to provide federated learning schemes with improved fairness. To tackle this challenge, we propose a novel federated learning system that employs zero-shot data augmentation on under-represented data to mitigate statistical heterogeneity and encourage more uniform accuracy performance across clients in federated networks. We study two variants of this scheme, Fed-ZDAC (federated learning with zero-shot data augmentation at the clients) and Fed-ZDAS (federated learning with zero-shot data augmentation at the server). Empirical results on a suite of datasets demonstrate the effectiveness of our methods on simultaneously improving the test accuracy and fairness.

Yoojin Choi, Jihwan Choi, Mostafa El-Khamy et al.

Network quantization is an essential procedure in deep learning for development of efficient fixed-point inference models on mobile or edge platforms. However, as datasets grow larger and privacy regulations become stricter, data sharing for model compression gets more difficult and restricted. In this paper, we consider data-free network quantization with synthetic data. The synthetic data are generated from a generator, while no data are used in training the generator and in quantization. To this end, we propose data-free adversarial knowledge distillation, which minimizes the maximum distance between the outputs of the teacher and the (quantized) student for any adversarial samples from a generator. To generate adversarial samples similar to the original data, we additionally propose matching statistics from the batch normalization layers for generated data and the original data in the teacher. Furthermore, we show the gain of producing diverse adversarial samples by using multiple generators and multiple students. Our experiments show the state-of-the-art data-free model compression and quantization results for (wide) residual networks and MobileNet on SVHN, CIFAR-10, CIFAR-100, and Tiny-ImageNet datasets. The accuracy losses compared to using the original datasets are shown to be very minimal.

Abdelrahman Abdelhamed, Mahmoud Afifi, Radu Timofte et al.

This paper reviews the NTIRE 2020 challenge on real image denoising with focus on the newly introduced dataset, the proposed methods and their results. The challenge is a new version of the previous NTIRE 2019 challenge on real image denoising that was based on the SIDD benchmark. This challenge is based on a newly collected validation and testing image datasets, and hence, named SIDD+. This challenge has two tracks for quantitatively evaluating image denoising performance in (1) the Bayer-pattern rawRGB and (2) the standard RGB (sRGB) color spaces. Each track ~250 registered participants. A total of 22 teams, proposing 24 methods, competed in the final phase of the challenge. The proposed methods by the participating teams represent the current state-of-the-art performance in image denoising targeting real noisy images. The newly collected SIDD+ datasets are publicly available at: https://bit.ly/siddplus_data.

Andreas Lugmayr, Martin Danelljan, Radu Timofte et al.

This paper reviews the NTIRE 2020 challenge on real world super-resolution. It focuses on the participating methods and final results. The challenge addresses the real world setting, where paired true high and low-resolution images are unavailable. For training, only one set of source input images is therefore provided along with a set of unpaired high-quality target images. In Track 1: Image Processing artifacts, the aim is to super-resolve images with synthetically generated image processing artifacts. This allows for quantitative benchmarking of the approaches \wrt a ground-truth image. In Track 2: Smartphone Images, real low-quality smart phone images have to be super-resolved. In both tracks, the ultimate goal is to achieve the best perceptual quality, evaluated using a human study. This is the second challenge on the subject, following AIM 2019, targeting to advance the state-of-the-art in super-resolution. To measure the performance we use the benchmark protocol from AIM 2019. In total 22 teams competed in the final testing phase, demonstrating new and innovative solutions to the problem.

Qingfeng Liu, Mostafa El-Khamy, Dongwoon Bai et al.

This paper proposes a novel deep learning architecture for semantic segmentation. The proposed Global and Selective Attention Network (GSANet) features Atrous Spatial Pyramid Pooling (ASPP) with a novel sparsemax global attention and a novel selective attention that deploys a condensation and diffusion mechanism to aggregate the multi-scale contextual information from the extracted deep features. A selective attention decoder is also proposed to process the GSA-ASPP outputs for optimizing the softmax volume. We are the first to benchmark the performance of semantic segmentation networks with the low-complexity feature extraction network (FXN) MobileNetEdge, that is optimized for low latency on edge devices. We show that GSANet can result in more accurate segmentation with MobileNetEdge, as well as with strong FXNs, such as Xception. GSANet improves the state-of-art semantic segmentation accuracy on both the ADE20k and the Cityscapes datasets.

Ryan Szeto, Mostafa El-Khamy, Jungwon Lee et al.

Video-to-video translation is more difficult than image-to-image translation due to the temporal consistency problem that, if unaddressed, leads to distracting flickering effects. Although video models designed from scratch produce temporally consistent results, training them to match the vast visual knowledge captured by image models requires an intractable number of videos. To combine the benefits of image and video models, we propose an image-to-video model transfer method called Hyperconsistency (HyperCon) that transforms any well-trained image model into a temporally consistent video model without fine-tuning. HyperCon works by translating a temporally interpolated video frame-wise and then aggregating over temporally localized windows on the interpolated video. It handles both masked and unmasked inputs, enabling support for even more video-to-video translation tasks than prior image-to-video model transfer techniques. We demonstrate HyperCon on video style transfer and inpainting, where it performs favorably compared to prior state-of-the-art methods without training on a single stylized or incomplete video. Our project website is available at https://ryanszeto.com/projects/hypercon .

Jaeyoung Kim, Mostafa El-Khamy, Jungwon Lee

Although supervised learning based on a deep neural network has recently achieved substantial improvement on speech enhancement, the existing schemes have either of two critical issues: spectrum or metric mismatches. The spectrum mismatch is a well known issue that any spectrum modification after short-time Fourier transform (STFT), in general, cannot be fully recovered after inverse short-time Fourier transform (ISTFT). The metric mismatch is that a conventional mean square error (MSE) loss function is typically sub-optimal to maximize perceptual speech measure such as signal-to-distortion ratio (SDR), perceptual evaluation of speech quality (PESQ) and short-time objective intelligibility (STOI). This paper presents a new end-to-end denoising framework. First, the network optimization is performed on the time-domain signals after ISTFT to avoid the spectrum mismatch. Second, three loss functions based on SDR, PESQ and STOI are proposed to minimize the metric mismatch. The experimental result showed the proposed denoising scheme significantly improved SDR, PESQ and STOI performance over the existing methods. Moreover, the proposed scheme also provided good generalization performance over generative denoising models on the perceptual speech metrics not used as a loss function during training.

Jaeyoung Kim, Mostafa El-Khamy, Jungwon Lee

Transformer neural networks (TNN) demonstrated state-of-art performance on many natural language processing (NLP) tasks, replacing recurrent neural networks (RNNs), such as LSTMs or GRUs. However, TNNs did not perform well in speech enhancement, whose contextual nature is different than NLP tasks, like machine translation. Self-attention is a core building block of the Transformer, which not only enables parallelization of sequence computation, but also provides the constant path length between symbols that is essential to learning long-range dependencies. In this paper, we propose a Transformer with Gaussian-weighted self-attention (T-GSA), whose attention weights are attenuated according to the distance between target and context symbols. The experimental results show that the proposed T-GSA has significantly improved speech-enhancement performance, compared to the Transformer and RNNs.

Yoojin Choi, Mostafa El-Khamy, Jungwon Lee

In this paper, we propose a novel variable-rate learned image compression framework with a conditional autoencoder. Previous learning-based image compression methods mostly require training separate networks for different compression rates so they can yield compressed images of varying quality. In contrast, we train and deploy only one variable-rate image compression network implemented with a conditional autoencoder. We provide two rate control parameters, i.e., the Lagrange multiplier and the quantization bin size, which are given as conditioning variables to the network. Coarse rate adaptation to a target is performed by changing the Lagrange multiplier, while the rate can be further fine-tuned by adjusting the bin size used in quantizing the encoded representation. Our experimental results show that the proposed scheme provides a better rate-distortion trade-off than the traditional variable-rate image compression codecs such as JPEG2000 and BPG. Our model also shows comparable and sometimes better performance than the state-of-the-art learned image compression models that deploy multiple networks trained for varying rates.

Mostafa El-Khamy, Haoyu Ren, Xianzhi Du et al.

In this paper, we introduce the problem of estimating the real world depth of elements in a scene captured by two cameras with different field of views, where the first field of view (FOV) is a Wide FOV (WFOV) captured by a wide angle lens, and the second FOV is contained in the first FOV and is captured by a tele zoom lens. We refer to the problem of estimating the inverse depth for the union of FOVs, while leveraging the stereo information in the overlapping FOV, as Tele-Wide Stereo Matching (TW-SM). We propose different deep learning solutions to the TW-SM problem. Since the disparity is proportional to the inverse depth, we train stereo matching disparity estimation (SMDE) networks to estimate the disparity for the union WFOV. We further propose an end-to-end deep multitask tele-wide stereo matching neural network (MT-TW-SMNet), which simultaneously learns the SMDE task for the overlapped Tele FOV and the single image inverse depth estimation (SIDE) task for the WFOV. Moreover, we design multiple methods for the fusion of the SMDE and SIDE networks. We evaluate the performance of TW-SM on the popular KITTI and SceneFlow stereo datasets, and demonstrate its practicality by synthesizing the Bokeh effect on the WFOV from a tele-wide stereo image pair.

Haoyu Ren, Mostafa El-khamy, Jungwon Lee

Single image depth estimation (SIDE) plays a crucial role in 3D computer vision. In this paper, we propose a two-stage robust SIDE framework that can perform blind SIDE for both indoor and outdoor scenes. At the first stage, the scene understanding module will categorize the RGB image into different depth-ranges. We introduce two different scene understanding modules based on scene classification and coarse depth estimation respectively. At the second stage, SIDE networks trained by the images of specific depth-range are applied to obtain an accurate depth map. In order to improve the accuracy, we further design a multi-task encoding-decoding SIDE network DS-SIDENet based on depthwise separable convolutions. DS-SIDENet is optimized to minimize both depth classification and depth regression losses. This improves the accuracy compared to a single-task SIDE network. Experimental results demonstrate that training DS-SIDENet on an individual dataset such as NYU achieves competitive performance to the state-of-art methods with much better efficiency. Ours proposed robust SIDE framework also shows good performance for the ScanNet indoor images and KITTI outdoor images simultaneously. It achieves the top performance compared to the Robust Vision Challenge (ROB) 2018 submissions.

J. Jon Ryu, Yoojin Choi, Young-Han Kim et al.

A new bimodal generative model is proposed for generating conditional and joint samples, accompanied with a training method with learning a succinct bottleneck representation. The proposed model, dubbed as the variational Wyner model, is designed based on two classical problems in network information theory -- distributed simulation and channel synthesis -- in which Wyner's common information arises as the fundamental limit on the succinctness of the common representation. The model is trained by minimizing the symmetric Kullback--Leibler divergence between variational and model distributions with regularization terms for common information, reconstruction consistency, and latent space matching terms, which is carried out via an adversarial density ratio estimation technique. The utility of the proposed approach is demonstrated through experiments for joint and conditional generation with synthetic and real-world datasets, as well as a challenging zero-shot image retrieval task.

Xianzhi Du, Mostafa El-Khamy, Jungwon Lee

In this paper, a new deep learning architecture for stereo disparity estimation is proposed. The proposed atrous multiscale network (AMNet) adopts an efficient feature extractor with depthwise-separable convolutions and an extended cost volume that deploys novel stereo matching costs on the deep features. A stacked atrous multiscale network is proposed to aggregate rich multiscale contextual information from the cost volume which allows for estimating the disparity with high accuracy at multiple scales. AMNet can be further modified to be a foreground-background aware network, FBA-AMNet, which is capable of discriminating between the foreground and the background objects in the scene at multiple scales. An iterative multitask learning method is proposed to train FBA-AMNet end-to-end. The proposed disparity estimation networks, AMNet and FBA-AMNet, show accurate disparity estimates and advance the state of the art on the challenging Middlebury, KITTI 2012, KITTI 2015, and Sceneflow stereo disparity estimation benchmarks.

Yoojin Choi, Mostafa El-Khamy, Jungwon Lee

We consider the optimization of deep convolutional neural networks (CNNs) such that they provide good performance while having reduced complexity if deployed on either conventional systems with spatial-domain convolution or lower-complexity systems designed for Winograd convolution. The proposed framework produces one compressed model whose convolutional filters can be made sparse either in the spatial domain or in the Winograd domain. Hence, the compressed model can be deployed universally on any platform, without need for re-training on the deployed platform. To get a better compression ratio, the sparse model is compressed in the spatial domain that has a fewer number of parameters. From our experiments, we obtain $24.2\times$ and $47.7\times$ compressed models for ResNet-18 and AlexNet trained on the ImageNet dataset, while their computational cost is also reduced by $4.5\times$ and $5.1\times$, respectively.

Jaeyoung Kim, Mostafa El-Khamy, Jungwon Lee

Supervised learning based on a deep neural network recently has achieved substantial improvement on speech enhancement. Denoising networks learn mapping from noisy speech to clean one directly, or to a spectrum mask which is the ratio between clean and noisy spectra. In either case, the network is optimized by minimizing mean square error (MSE) between ground-truth labels and time-domain or spectrum output. However, existing schemes have either of two critical issues: spectrum and metric mismatches. The spectrum mismatch is a well known issue that any spectrum modification after short-time Fourier transform (STFT), in general, cannot be fully recovered after inverse short-time Fourier transform (ISTFT). The metric mismatch is that a conventional MSE metric is sub-optimal to maximize our target metrics, signal-to-distortion ratio (SDR) and perceptual evaluation of speech quality (PESQ). This paper presents a new end-to-end denoising framework with the goal of joint SDR and PESQ optimization. First, the network optimization is performed on the time-domain signals after ISTFT to avoid spectrum mismatch. Second, two loss functions which have improved correlations with SDR and PESQ metrics are proposed to minimize metric mismatch. The experimental result showed that the proposed denoising scheme significantly improved both SDR and PESQ performance over the existing methods.

Haoyu Ren, Mostafa El-Khamy, Jungwon Lee

A deep learning approach to blind denoising of images without complete knowledge of the noise statistics is considered. We propose DN-ResNet, which is a deep convolutional neural network (CNN) consisting of several residual blocks (ResBlocks). With cascade training, DN-ResNet is more accurate and more computationally efficient than the state of art denoising networks. An edge-aware loss function is further utilized in training DN-ResNet, so that the denoising results have better perceptive quality compared to conventional loss function. Next, we introduce the depthwise separable DN-ResNet (DS-DN-ResNet) utilizing the proposed Depthwise Seperable ResBlock (DS-ResBlock) instead of standard ResBlock, which has much less computational cost. DS-DN-ResNet is incrementally evolved by replacing the ResBlocks in DN-ResNet by DS-ResBlocks stage by stage. As a result, high accuracy and good computational efficiency are achieved concurrently. Whereas previous state of art deep learning methods focused on denoising either Gaussian or Poisson corrupted images, we consider denoising images having the more practical Poisson with additive Gaussian noise as well. The results show that DN-ResNets are more efficient, robust, and perform better denoising than current state of art deep learning methods, as well as the popular variants of the BM3D algorithm, in cases of blind and non-blind denoising of images corrupted with Poisson, Gaussian or Poisson-Gaussian noise. Our network also works well for other image enhancement task such as compressed image restoration.

Yoojin Choi, Mostafa El-Khamy, Jungwon Lee

We consider learning deep neural networks (DNNs) that consist of low-precision weights and activations for efficient inference of fixed-point operations. In training low-precision networks, gradient descent in the backward pass is performed with high-precision weights while quantized low-precision weights and activations are used in the forward pass to calculate the loss function for training. Thus, the gradient descent becomes suboptimal, and accuracy loss follows. In order to reduce the mismatch in the forward and backward passes, we utilize mean squared quantization error (MSQE) regularization. In particular, we propose using a learnable regularization coefficient with the MSQE regularizer to reinforce the convergence of high-precision weights to their quantized values. We also investigate how partial L2 regularization can be employed for weight pruning in a similar manner. Finally, combining weight pruning, quantization, and entropy coding, we establish a low-precision DNN compression pipeline. In our experiments, the proposed method yields low-precision MobileNet and ShuffleNet models on ImageNet classification with the state-of-the-art compression ratios of 7.13 and 6.79, respectively. Moreover, we examine our method for image super resolution networks to produce 8-bit low-precision models at negligible performance loss.

Yoojin Choi, Mostafa El-Khamy, Jungwon Lee

We consider the optimization of deep convolutional neural networks (CNNs) such that they provide good performance while having reduced complexity if deployed on either conventional systems utilizing spatial-domain convolution or lower complexity systems designed for Winograd convolution. Furthermore, we explore the universal quantization and compression of these networks. In particular, the proposed framework produces one compressed model whose convolutional filters can be made sparse either in the spatial domain or in the Winograd domain. Hence, one compressed model can be deployed universally on any platform, without need for re-training on the deployed platform, and the sparsity of its convolutional filters can be exploited for further complexity reduction in either domain. To get a better compression ratio, the sparse model is compressed in the spatial domain which has a less number of parameters. From our experiments, we obtain $24.2\times$, $47.7\times$ and $35.4\times$ compressed models for ResNet-18, AlexNet and CT-SRCNN, while their computational cost is also reduced by $4.5\times$, $5.1\times$ and $23.5\times$, respectively.

Xianzhi Du, Mostafa El-Khamy, Vlad I. Morariu et al.

In this paper, we present an efficient pedestrian detection system, designed by fusion of multiple deep neural network (DNN) systems. Pedestrian candidates are first generated by a single shot convolutional multi-box detector at different locations with various scales and aspect ratios. The candidate generator is designed to provide the majority of ground truth pedestrian annotations at the cost of a large number of false positives. Then, a classification system using the idea of ensemble learning is deployed to improve the detection accuracy. The classification system further classifies the generated candidates based on opinions of multiple deep verification networks and a fusion network which utilizes a novel soft-rejection fusion method to adjust the confidence in the detection results. To improve the training of the deep verification networks, a novel soft-label method is devised to assign floating point labels to the generated pedestrian candidates. A deep context aggregation semantic segmentation network also provides pixel-level classification of the scene and its results are softly fused with the detection results by the single shot detector. Our pedestrian detector compared favorably to state-of-art methods on all popular pedestrian detection datasets. For example, our fused DNN has better detection accuracy on the Caltech Pedestrian dataset than all previous state of art methods, while also being the fastest. We significantly improved the log-average miss rate on the Caltech pedestrian dataset to 7.67% and achieved the new state-of-the-art.

Yoojin Choi, Mostafa El-Khamy, Jungwon Lee

In this paper, we investigate lossy compression of deep neural networks (DNNs) by weight quantization and lossless source coding for memory-efficient deployment. Whereas the previous work addressed non-universal scalar quantization and entropy coding of DNN weights, we for the first time introduce universal DNN compression by universal vector quantization and universal source coding. In particular, we examine universal randomized lattice quantization of DNNs, which randomizes DNN weights by uniform random dithering before lattice quantization and can perform near-optimally on any source without relying on knowledge of its probability distribution. Moreover, we present a method of fine-tuning vector quantized DNNs to recover the performance loss after quantization. Our experimental results show that the proposed universal DNN compression scheme compresses the 32-layer ResNet (trained on CIFAR-10) and the AlexNet (trained on ImageNet) with compression ratios of $47.1$ and $42.5$, respectively.

Haoyu Ren, Mostafa El-Khamy, Jungwon Lee

We propose methodologies to train highly accurate and efficient deep convolutional neural networks (CNNs) for image super resolution (SR). A cascade training approach to deep learning is proposed to improve the accuracy of the neural networks while gradually increasing the number of network layers. Next, we explore how to improve the SR efficiency by making the network slimmer. Two methodologies, the one-shot trimming and the cascade trimming, are proposed. With the cascade trimming, the network's size is gradually reduced layer by layer, without significant loss on its discriminative ability. Experiments on benchmark image datasets show that our proposed SR network achieves the state-of-the-art super resolution accuracy, while being more than 4 times faster compared to existing deep super resolution networks.

Jaeyoung Kim, Mostafa El-Khamy, Jungwon Lee

Despite the remarkable progress achieved on automatic speech recognition, recognizing far-field speeches mixed with various noise sources is still a challenging task. In this paper, we introduce novel student-teacher transfer learning, BridgeNet which can provide a solution to improve distant speech recognition. There are two key features in BridgeNet. First, BridgeNet extends traditional student-teacher frameworks by providing multiple hints from a teacher network. Hints are not limited to the soft labels from a teacher network. Teacher's intermediate feature representations can better guide a student network to learn how to denoise or dereverberate noisy input. Second, the proposed recursive architecture in the BridgeNet can iteratively improve denoising and recognition performance. The experimental results of BridgeNet showed significant improvements in tackling the distant speech recognition problem, where it achieved up to 13.24% relative WER reductions on AMI corpus compared to a baseline neural network without teacher's hints.

Jaeyoung Kim, Mostafa El-Khamy, Jungwon Lee

In this paper, a novel architecture for a deep recurrent neural network, residual LSTM is introduced. A plain LSTM has an internal memory cell that can learn long term dependencies of sequential data. It also provides a temporal shortcut path to avoid vanishing or exploding gradients in the temporal domain. The residual LSTM provides an additional spatial shortcut path from lower layers for efficient training of deep networks with multiple LSTM layers. Compared with the previous work, highway LSTM, residual LSTM separates a spatial shortcut path with temporal one by using output layers, which can help to avoid a conflict between spatial and temporal-domain gradient flows. Furthermore, residual LSTM reuses the output projection matrix and the output gate of LSTM to control the spatial information flow instead of additional gate networks, which effectively reduces more than 10% of network parameters. An experiment for distant speech recognition on the AMI SDM corpus shows that 10-layer plain and highway LSTM networks presented 13.7% and 6.2% increase in WER over 3-layer aselines, respectively. On the contrary, 10-layer residual LSTM networks provided the lowest WER 41.0%, which corresponds to 3.3% and 2.8% WER reduction over plain and highway LSTM networks, respectively.

Yoojin Choi, Mostafa El-Khamy, Jungwon Lee

Network quantization is one of network compression techniques to reduce the redundancy of deep neural networks. It reduces the number of distinct network parameter values by quantization in order to save the storage for them. In this paper, we design network quantization schemes that minimize the performance loss due to quantization given a compression ratio constraint. We analyze the quantitative relation of quantization errors to the neural network loss function and identify that the Hessian-weighted distortion measure is locally the right objective function for the optimization of network quantization. As a result, Hessian-weighted k-means clustering is proposed for clustering network parameters to quantize. When optimal variable-length binary codes, e.g., Huffman codes, are employed for further compression, we derive that the network quantization problem can be related to the entropy-constrained scalar quantization (ECSQ) problem in information theory and consequently propose two solutions of ECSQ for network quantization, i.e., uniform quantization and an iterative solution similar to Lloyd's algorithm. Finally, using the simple uniform quantization followed by Huffman coding, we show from our experiments that the compression ratios of 51.25, 22.17 and 40.65 are achievable for LeNet, 32-layer ResNet and AlexNet, respectively.

Xianzhi Du, Mostafa El-Khamy, Jungwon Lee et al.

We propose a deep neural network fusion architecture for fast and robust pedestrian detection. The proposed network fusion architecture allows for parallel processing of multiple networks for speed. A single shot deep convolutional network is trained as a object detector to generate all possible pedestrian candidates of different sizes and occlusions. This network outputs a large variety of pedestrian candidates to cover the majority of ground-truth pedestrians while also introducing a large number of false positives. Next, multiple deep neural networks are used in parallel for further refinement of these pedestrian candidates. We introduce a soft-rejection based network fusion method to fuse the soft metrics from all networks together to generate the final confidence scores. Our method performs better than existing state-of-the-arts, especially when detecting small-size and occluded pedestrians. Furthermore, we propose a method for integrating pixel-wise semantic segmentation network into the network fusion architecture as a reinforcement to the pedestrian detector. The approach outperforms state-of-the-art methods on most protocols on Caltech Pedestrian dataset, with significant boosts on several protocols. It is also faster than all other methods.