Calcium imaging is an essential tool to study the activity of neuronal populations. However, the high level of background fluorescence in images hinders the accurate identification of neurons and the extraction of neuronal activities. While robust principal component analysis (RPCA) is a promising method that can decompose the foreground and background in such images, its computational complexity and memory requirement are prohibitively high to process large-scale calcium imaging data. Here, we propose BEAR, a simple bilinear neural network for the efficient approximation of RPCA which achieves an order of magnitude speed improvement with GPU acceleration compared to the conventional RPCA algorithms. In addition, we show that BEAR can perform foreground-background separation of calcium imaging data as large as tens of gigabytes. We also demonstrate that two BEARs can be cascaded to perform simultaneous RPCA and non-negative matrix factorization for the automated extraction of spatial and temporal footprints from calcium imaging data. The source code used in the paper is available at https://github.com/NICALab/BEAR.
Graph layouts are key to exploring massive graphs. An enormous number of nodes and edges do not allow network analysis software to produce meaningful visualization of the pervasive networks. Long computation time, memory and display limitations encircle the software's ability to explore massive graphs. This paper introduces BigGraphVis, a new parallel graph visualization method that uses GPU parallel processing and community detection algorithm to visualize graph communities. We combine parallelized streaming community detection algorithm and probabilistic data structure to leverage parallel processing of Graphics Processing Unit (GPU). To the best of our knowledge, this is the first attempt to combine the power of streaming algorithms coupled with GPU computing to tackle big graph visualization challenges. Our method extracts community information in a few passes on the edge list, and renders the community structures using the ForceAtlas2 algorithm. Our experiment with massive real-life graphs indicates that about 70 to 95 percent speedup can be achieved by visualizing graph communities, and the visualization appears to be meaningful and reliable. The biggest graph that we examined contains above 3 million nodes and 34 million edges, and the layout computation took about five minutes. We also observed that the BigGraphVis coloring strategy can be successfully applied to produce a more informative ForceAtlas2 layout.
Statistical machine learning has widespread application in various domains. These methods include probabilistic algorithms, such as Markov Chain Monte-Carlo (MCMC), which rely on generating random numbers from probability distributions. These algorithms are computationally expensive on conventional processors, yet their statistical properties, namely interpretability and uncertainty quantification (UQ) compared to deep learning, make them an attractive alternative approach. Therefore, hardware specialization can be adopted to address the shortcomings of conventional processors in running these applications. In this paper, we propose a high-throughput accelerator for Markov Random Field (MRF) inference, a powerful model for representing a wide range of applications, using MCMC with Gibbs sampling. We propose a tiled architecture which takes advantage of near-memory computing, and memory optimizations tailored to the semantics of MRF. Additionally, we propose a novel hybrid on-chip/off-chip memory system and logging scheme to efficiently support UQ. This memory system design is not specific to MRF models and is applicable to applications using probabilistic algorithms. In addition, it dramatically reduces off-chip memory bandwidth requirements. We implemented an FPGA prototype of our proposed architecture using high-level synthesis tools and achieved 146MHz frequency for an accelerator with 32 function units on an Intel Arria 10 FPGA. Compared to prior work on FPGA, our accelerator achieves 26X speedup. Furthermore, our proposed memory system and logging scheme to support UQ reduces off-chip bandwidth by 71% for two applications. ASIC analysis in 15nm shows our design with 2048 function units running at 3GHz outperforms GPU implementations of motion estimation and stereo vision on Nvidia RTX2080Ti by 120X-210X, occupying only 7.7% of the area.
Latest CNN-based object detection models are quite accurate but require a high-performance GPU to run in real-time. They still are heavy in terms of memory size and speed for an embedded system with limited memory space. Since the object detection for autonomous system is run on an embedded processor, it is preferable to compress the detection network as light as possible while preserving the detection accuracy. There are several popular lightweight detection models but their accuracy is too low for safe driving applications. Therefore, this paper proposes a new object detection model, referred as YOffleNet, which is compressed at a high ratio while minimizing the accuracy loss for real-time and safe driving application on an autonomous system. The backbone network architecture is based on YOLOv4, but we could compress the network greatly by replacing the high-calculation-load CSP DenseNet with the lighter modules of ShuffleNet. Experiments with KITTI dataset showed that the proposed YOffleNet is compressed by 4.7 times than the YOLOv4-s that could achieve as fast as 46 FPS on an embedded GPU system(NVIDIA Jetson AGX Xavier). Compared to the high compression ratio, the accuracy is reduced slightly to 85.8% mAP, that is only 2.6% lower than YOLOv4-s. Thus, the proposed network showed a high potential to be deployed on the embedded system of the autonomous system for the real-time and accurate object detection applications.
Facial landmarks (FLM) estimation is a critical component in many face-related applications. In this work, we aim to optimize for both accuracy and speed and explore the trade-off between them. Our key observation is that not all faces are created equal. Frontal faces with neutral expressions converge faster than faces with extreme poses or expressions. To differentiate among samples, we train our model to predict the regression error after each iteration. If the current iteration is accurate enough, we stop iterating, saving redundant iterations while keeping the accuracy in check. We also observe that as neighboring patches overlap, we can infer all facial landmarks (FLMs) with only a small number of patches without a major accuracy sacrifice. Architecturally, we offer a multi-scale, patch-based, lightweight feature extractor with a fine-grained local patch attention module, which computes a patch weighting according to the information in the patch itself and enhances the expressive power of the patch features. We analyze the patch attention data to infer where the model is attending when regressing facial landmarks and compare it to face attention in humans. Our model runs in real-time on a mobile device GPU, with 95 Mega Multiply-Add (MMA) operations, outperforming all state-of-the-art methods under 1000 MMA, with a normalized mean error of 8.16 on the 300W challenging dataset.
Recently, Information Retrieval community has witnessed fast-paced advances in Dense Retrieval (DR), which performs first-stage retrieval with embedding-based search. Despite the impressive ranking performance, previous studies usually adopt brute-force search to acquire candidates, which is prohibitive in practical Web search scenarios due to its tremendous memory usage and time cost. To overcome these problems, vector compression methods have been adopted in many practical embedding-based retrieval applications. One of the most popular methods is Product Quantization (PQ). However, although existing vector compression methods including PQ can help improve the efficiency of DR, they incur severely decayed retrieval performance due to the separation between encoding and compression. To tackle this problem, we present JPQ, which stands for Joint optimization of query encoding and Product Quantization. It trains the query encoder and PQ index jointly in an end-to-end manner based on three optimization strategies, namely ranking-oriented loss, PQ centroid optimization, and end-to-end negative sampling. We evaluate JPQ on two publicly available retrieval benchmarks. Experimental results show that JPQ significantly outperforms popular vector compression methods. Compared with previous DR models that use brute-force search, JPQ almost matches the best retrieval performance with 30x compression on index size. The compressed index further brings 10x speedup on CPU and 2x speedup on GPU in query latency.
Video based fall detection accuracy has been largely improved due to the recent progress on deep convolutional neural networks. However, there still exists some challenges, such as lighting variation, complex background, which degrade the accuracy and generalization ability of these approaches. Meanwhile, large computation cost limits the application of existing fall detection approaches. To alleviate these problems, a video based fall detection approach using human poses is proposed in this paper. First, a lightweight pose estimator extracts 2D poses from video sequences and then 2D poses are lifted to 3D poses. Second, we introduce a robust fall detection network to recognize fall events using estimated 3D poses, which increases respective filed and maintains low computation cost by dilated convolutions. The experimental results show that the proposed fall detection approach achieves a high accuracy of 99.83% on large benchmark action recognition dataset NTU RGB+D and real-time performance of 18 FPS on a non-GPU platform and 63 FPS on a GPU platform.
Recent advances in deep learning have drastically improved performance on many Natural Language Understanding (NLU) tasks. However, the data used to train NLU models may contain private information such as addresses or phone numbers, particularly when drawn from human subjects. It is desirable that underlying models do not expose private information contained in the training data. Differentially Private Stochastic Gradient Descent (DP-SGD) has been proposed as a mechanism to build privacy-preserving models. However, DP-SGD can be prohibitively slow to train. In this work, we propose a more efficient DP-SGD for training using a GPU infrastructure and apply it to fine-tuning models based on LSTM and transformer architectures. We report faster training times, alongside accuracy, theoretical privacy guarantees and success of Membership inference attacks for our models and observe that fine-tuning with proposed variant of DP-SGD can yield competitive models without significant degradation in training time and improvement in privacy protection. We also make observations such as looser theoretical $\epsilon, \delta$ can translate into significant practical privacy gains.
We present an out-of-core variational approach for surface reconstruction from a set of aligned depth maps. Input depth maps are supposed to be reconstructed from regular photos or/and can be a representation of terrestrial LIDAR point clouds. Our approach is based on surface reconstruction via total generalized variation minimization ($TGV$) because of its strong visibility-based noise-filtering properties and GPU-friendliness. Our main contribution is an out-of-core OpenCL-accelerated adaptation of this numerical algorithm which can handle arbitrarily large real-world scenes with scale diversity.
In this report, we introduce our winning solution to the Real-time 3D Detection and also the "Most Efficient Model" in the Waymo Open Dataset Challenges at CVPR 2021. Extended from our last year's award-winning model AFDet, we have made a handful of modifications to the base model, to improve the accuracy and at the same time to greatly reduce the latency. The modified model, named as AFDetV2, is featured with a lite 3D Feature Extractor, an improved RPN with extended receptive field and an added sub-head that produces an IoU-aware confidence score. These model enhancements, together with enriched data augmentation, stochastic weights averaging, and a GPU-based implementation of voxelization, lead to a winning accuracy of 73.12 mAPH/L2 for our AFDetV2 with a latency of 60.06 ms, and an accuracy of 72.57 mAPH/L2 for our AFDetV2-base, entitled as the "Most Efficient Model" by the challenge sponsor, with a winning latency of 55.86 ms.
Deep learning workloads are common in today’s production clusters due to the proliferation of deep learning-driven AI services (e.g., speech recognition, machine translation). A deep learning training job is resource-intensive and time-consuming. Efficient resource scheduling is the key to the maximal performance of a deep learning cluster. Existing cluster schedulers are largely not tailored to deep learning jobs, and typically specifying a fixed amount of resources for each job, prohibiting high resource efficiency and job performance. This paper proposes Optimus, a customized job scheduler for deep learning clusters, which minimizes job training time based on online resource-performance models. Optimus uses online fitting to predict model convergence during training and sets up performance models to accurately estimate training speed as a function of allocated resources in each job. Based on the models, a simple yet effective method is designed and used for dynamically allocating resources and placing deep learning tasks to minimize job completion time. We implement Optimus on top of Kubernetes, a cluster manager for container orchestration, and experiment on a deep learning cluster with 7 CPU servers and 6 GPU servers, running 9 training jobs using the MXNet framework. Results show that Optimus outperforms representative cluster schedulers by about 139% and 63% in terms of job completion time and makespan, respectively
In recent years, federated learning (FL) has been widely applied for supporting decentralized collaborative learning scenarios. Among existing FL models, federated logistic regression (FLR) is a widely used statistic model and has been used in various industries. To ensure data security and user privacy, FLR leverages homomorphic encryption (HE) to protect the exchanged data among different collaborative parties. However, HE introduces significant computational overhead (i.e., the cost of data encryption/decryption and calculation over encrypted data), which eventually becomes the performance bottleneck of the whole system. In this paper, we propose HAFLO, a GPU-based solution to improve the performance of FLR. The core idea of HAFLO is to summarize a set of performance-critical homomorphic operators (HO) used by FLR and accelerate the execution of these operators through a joint optimization of storage, IO, and computation. The preliminary results show that our acceleration on FATE, a popular FL framework, achieves a 49.9$\times$ speedup for heterogeneous LR and 88.4$\times$ for homogeneous LR.
3D time-of-flight (ToF) imaging is used in a variety of applications such as augmented reality (AR), computer interfaces, robotics and autonomous systems. Single-photon avalanche diodes (SPADs) are one of the enabling technologies providing accurate depth data even over long ranges. By developing SPADs in array format with integrated processing combined with pulsed, flood-type illumination, high-speed 3D capture is possible. However, array sizes tend to be relatively small, limiting the lateral resolution of the resulting depth maps, and, consequently, the information that can be extracted from the image for applications such as object detection. In this paper, we demonstrate that these limitations can be overcome through the use of convolutional neural networks (CNNs) for high-performance object detection. We present outdoor results from a portable SPAD camera system that outputs 16-bin photon timing histograms with 64x32 spatial resolution. The results, obtained with exposure times down to 2 ms (equivalent to 500 FPS) and in signal-to-background (SBR) ratios as low as 0.05, point to the advantages of providing the CNN with full histogram data rather than point clouds alone. Alternatively, a combination of point cloud and active intensity data may be used as input, for a similar level of performance. In either case, the GPU-accelerated processing time is less than 1 ms per frame, leading to an overall latency (image acquisition plus processing) in the millisecond range, making the results relevant for safety-critical computer vision applications which would benefit from faster than human reaction times.
As the proportion of road accidents increases each year, driver distraction continues to be an important risk component in road traffic injuries and deaths. The distractions caused by the increasing use of mobile phones and other wireless devices pose a potential risk to road safety. Our current study aims to aid the already existing techniques in driver posture recognition by improving the performance in the driver distraction classification problem. We present an approach using a genetic algorithm-based ensemble of six independent deep neural architectures, namely, AlexNet, VGG-16, EfficientNet B0, Vanilla CNN, Modified DenseNet, and InceptionV3 + BiLSTM. We test it on two comprehensive datasets, the AUC Distracted Driver Dataset, on which our technique achieves an accuracy of 96.37%, surpassing the previously obtained 95.98%, and on the State Farm Driver Distraction Dataset, on which we attain an accuracy of 99.75%. The 6-Model Ensemble gave an inference time of 0.024 seconds as measured on our machine with Ubuntu 20.04(64-bit) and GPU as GeForce GTX 1080.
Although spiking neural networks (SNNs) take benefits from the bio-plausible neural modeling, the low accuracy under the common local synaptic plasticity learning rules limits their application in many practical tasks. Recently, an emerging SNN supervised learning algorithm inspired by backpropagation through time (BPTT) from the domain of artificial neural networks (ANNs) has successfully boosted the accuracy of SNNs and helped improve the practicability of SNNs. However, current general-purpose processors suffer from low efficiency when performing BPTT for SNNs due to the ANN-tailored optimization. On the other hand, current neuromorphic chips cannot support BPTT because they mainly adopt local synaptic plasticity rules for simplified implementation. In this work, we propose H2Learn, a novel architecture that can achieve high efficiency for BPTT-based SNN learning which ensures high accuracy of SNNs. At the beginning, we characterized the behaviors of BPTT-based SNN learning. Benefited from the binary spike-based computation in the forward pass and the weight update, we first design lookup table (LUT) based processing elements in Forward Engine and Weight Update Engine to make accumulations implicit and to fuse the computations of multiple input points. Second, benefited from the rich sparsity in the backward pass, we design a dual-sparsity-aware Backward Engine which exploits both input and output sparsity. Finally, we apply a pipeline optimization between different engines to build an end-to-end solution for the BPTT-based SNN learning. Compared with the modern NVIDIA V100 GPU, H2Learn achieves 7.38x area saving, 5.74-10.20x speedup, and 5.25-7.12x energy saving on several benchmark datasets.
Classifiers learnt from data are increasingly being used as components in systems where safety is a critical concern. In this work, we present a formal notion of safety for classifiers via constraints called safe-ordering constraints. These constraints relate requirements on the order of the classes output by a classifier to conditions on its input, and are expressive enough to encode various interesting examples of classifier safety specifications from the literature. For classifiers implemented using neural networks, we also present a run-time mechanism for the enforcement of safe-ordering constraints. Our approach is based on a self-correcting layer, which provably yields safe outputs regardless of the characteristics of the classifier input. We compose this layer with an existing neural network classifier to construct a self-correcting network (SC-Net), and show that in addition to providing safe outputs, the SC-Net is guaranteed to preserve the classification accuracy of the original network whenever possible. Our approach is independent of the size and architecture of the neural network used for classification, depending only on the specified property and the dimension of the network's output; thus it is scalable to large state-of-the-art networks. We show that our approach can be optimized for a GPU, introducing run-time overhead of less than 1ms on current hardware -- even on large, widely-used networks containing hundreds of thousands of neurons and millions of parameters.
Mobile devices are transforming the way people interact with computers, and speech interfaces to applications are ever more important. Automatic Speech Recognition systems recently published are very accurate, but often require powerful machinery (specialised Graphical Processing Units) for inference, which makes them impractical to run on commodity devices, especially in streaming mode. Impressed by the accuracy of, but dissatisfied with the inference times of the baseline Kazakh ASR model of (Khassanov et al.,2021) when not using a GPU, we trained a new baseline acoustic model (on the same dataset as the aforementioned paper) and three language models for use with the Coqui STT framework. Results look promising, but further epochs of training and parameter sweeping or, alternatively, limiting the vocabulary that the ASR system must support, is needed to reach a production-level accuracy.
We introduce latency-aware network acceleration (LANA) - an approach that builds on neural architecture search techniques and teacher-student distillation to accelerate neural networks. LANA consists of two phases: in the first phase, it trains many alternative operations for every layer of the teacher network using layer-wise feature map distillation. In the second phase, it solves the combinatorial selection of efficient operations using a novel constrained integer linear optimization (ILP) approach. ILP brings unique properties as it (i) performs NAS within a few seconds to minutes, (ii) easily satisfies budget constraints, (iii) works on the layer-granularity, (iv) supports a huge search space $O(10^{100})$, surpassing prior search approaches in efficacy and efficiency. In extensive experiments, we show that LANA yields efficient and accurate models constrained by a target latency budget, while being significantly faster than other techniques. We analyze three popular network architectures: EfficientNetV1, EfficientNetV2 and ResNeST, and achieve accuracy improvement for all models (up to $3.0\%$) when compressing larger models to the latency level of smaller models. LANA achieves significant speed-ups (up to $5\times$) with minor to no accuracy drop on GPU and CPU. The code will be shared soon.
In surveillance and search and rescue applications, it is important to perform multi-target tracking (MOT) in real-time on low-end devices. Today's MOT solutions employ deep neural networks, which tend to have high computation complexity. Recognizing the effects of frame sizes on tracking performance, we propose DeepScale, a model agnostic frame size selection approach that operates on top of existing fully convolutional network-based trackers to accelerate tracking throughput. In the training stage, we incorporate detectability scores into a one-shot tracker architecture so that DeepScale can learn representation estimations for different frame sizes in a self-supervised manner. During inference, it can adapt frame sizes according to the complexity of visual contents based on user-controlled parameters. To leverage computation resources on edge servers, we propose two computation partition schemes tailored for MOT, namely, edge server only with adaptive frame-size transmission and edge server-assisted tracking. Extensive experiments and benchmark tests on MOT datasets demonstrate the effectiveness and flexibility of DeepScale. Compared to a state-of-the-art tracker, DeepScale++, a variant of DeepScale achieves 1.57X accelerated with only moderate degradation ~2.3\ in tracking accuracy on the MOT15 dataset in one configuration. We have implemented and evaluated DeepScale++ and the proposed computation partition schemes on a small-scale testbed consisting of an NVIDIA Jetson TX2 board and a GPU server. The experiments reveal non-trivial trade-offs between tracking performance and latency compared to server-only or smart camera-only solutions.
We present DrQ-v2, a model-free reinforcement learning (RL) algorithm for visual continuous control. DrQ-v2 builds on DrQ, an off-policy actor-critic approach that uses data augmentation to learn directly from pixels. We introduce several improvements that yield state-of-the-art results on the DeepMind Control Suite. Notably, DrQ-v2 is able to solve complex humanoid locomotion tasks directly from pixel observations, previously unattained by model-free RL. DrQ-v2 is conceptually simple, easy to implement, and provides significantly better computational footprint compared to prior work, with the majority of tasks taking just 8 hours to train on a single GPU. Finally, we publicly release DrQ-v2's implementation to provide RL practitioners with a strong and computationally efficient baseline.