Pairwise sequence alignment is one of the most computationally intensive kernels in genomic data analysis, accounting for more than 90% of the runtime for key bioinformatics applications. This method is particularly expensive for third-generation sequences due to the high computational cost of analyzing sequences of length between 1Kb and 1Mb. Given the quadratic overhead of exact pairwise algorithms for long alignments, the community primarily relies on approximate algorithms that search only for high-quality alignments and stop early when one is not found. In this work, we present the first GPU optimization of the popular X-drop alignment algorithm, that we named LOGAN. Results show that our high-performance multi-GPU implementation achieves up to 181.6 GCUPS and speed-ups up to 6.6x and 30.7x using 1 and 6 NVIDIA Tesla V100, respectively, over the state-of-the-art software running on two IBM Power9 processors using 168 CPU threads, with equivalent accuracy. We also demonstrate a 2.3x LOGAN speed-up versus ksw2, a state-of-art vectorized algorithm for sequence alignment implemented in minimap2, a long-read mapping software. To highlight the impact of our work on a real-world application, we couple LOGAN with a many-to-many long-read alignment software called BELLA, and demonstrate that our implementation improves the overall BELLA runtime by up to 10.6x. Finally, we adapt the Roofline model for LOGAN and demonstrate that our implementation is near-optimal on the NVIDIA Tesla V100s.
Building conversational systems in new domains and with added functionality requires resource-efficient models that work under low-data regimes (i.e., in few-shot setups). Motivated by these requirements, we introduce intent detection methods backed by pretrained dual sentence encoders such as USE and ConveRT. We demonstrate the usefulness and wide applicability of the proposed intent detectors, showing that: 1) they outperform intent detectors based on fine-tuning the full BERT-Large model or using BERT as a fixed black-box encoder on three diverse intent detection data sets; 2) the gains are especially pronounced in few-shot setups (i.e., with only 10 or 30 annotated examples per intent); 3) our intent detectors can be trained in a matter of minutes on a single CPU; and 4) they are stable across different hyperparameter settings. In hope of facilitating and democratizing research focused on intention detection, we release our code, as well as a new challenging single-domain intent detection dataset comprising 13,083 annotated examples over 77 intents.
Accurate motion state estimation of Vulnerable Road Users (VRUs), is a critical requirement for autonomous vehicles that navigate in urban environments. Due to their computational efficiency, many traditional autonomy systems perform multi-object tracking using Kalman Filters which frequently rely on hand-engineered association. However, such methods fail to generalize to crowded scenes and multi-sensor modalities, often resulting in poor state estimates which cascade to inaccurate predictions. We present a practical and lightweight tracking system, SDVTracker, that uses a deep learned model for association and state estimation in conjunction with an Interacting Multiple Model (IMM) filter. The proposed tracking method is fast, robust and generalizes across multiple sensor modalities and different VRU classes. In this paper, we detail a model that jointly optimizes both association and state estimation with a novel loss, an algorithm for determining ground-truth supervision, and a training procedure. We show this system significantly outperforms hand-engineered methods on a real-world urban driving dataset while running in less than 2.5 ms on CPU for a scene with 100 actors, making it suitable for self-driving applications where low latency and high accuracy is critical.
This article presents an automatic approach to quickly derive a good solution for hardware resource partition and task granularity for task-based parallel applications on heterogeneous many-core architectures. Our approach employs a performance model to estimate the resulting performance of the target application under a given resource partition and task granularity configuration. The model is used as a utility to quickly search for a good configuration at runtime. Instead of hand-crafting an analytical model that requires expert insights into low-level hardware details, we employ machine learning techniques to automatically learn it. We achieve this by first learning a predictive model offline using training programs. The learnt model can then be used to predict the performance of any unseen program at runtime. We apply our approach to 39 representative parallel applications and evaluate it on two representative heterogeneous many-core platforms: a CPU-XeonPhi platform and a CPU-GPU platform. Compared to the single-stream version, our approach achieves, on average, a 1.6x and 1.1x speedup on the XeonPhi and the GPU platform, respectively. These results translate to over 93% of the performance delivered by a theoretically perfect predictor.
Recently, heatmap regression models have become popular due to their superior performance in locating facial landmarks. However, three major problems still exist among these models: (1) they are computationally expensive; (2) they usually lack explicit constraints on global shapes; (3) domain gaps are commonly present. To address these problems, we propose Pixel-in-Pixel Net (PIPNet) for facial landmark detection. The proposed model is equipped with a novel detection head based on heatmap regression, which conducts score and offset predictions simultaneously on low-resolution feature maps. By doing so, repeated upsampling layers are no longer necessary, enabling the inference time to be largely reduced without sacrificing model accuracy. Besides, a simple but effective neighbor regression module is proposed to enforce local constraints by fusing predictions from neighboring landmarks, which enhances the robustness of the new detection head. To further improve the cross-domain generalization capability of PIPNet, we propose self-training with curriculum. This training strategy is able to mine more reliable pseudo-labels from unlabeled data across domains by starting with an easier task, then gradually increasing the difficulty to provide more precise labels. Extensive experiments demonstrate the superiority of PIPNet, which obtains state-of-the-art results on three out of six popular benchmarks under the supervised setting. The results on two cross-domain test sets are also consistently improved compared to the baselines. Notably, our lightweight version of PIPNet runs at 35.7 FPS and 200 FPS on CPU and GPU, respectively, while still maintaining a competitive accuracy to state-of-the-art methods. The code of PIPNet is available at https://github.com/jhb86253817/PIPNet.
Graph processors such as Graphcore's Intelligence Processing Unit (IPU) are part of the major new wave of novel computer architecture for AI, and have a general design with massively parallel computation, distributed on-chip memory and very high inter-core communication bandwidth which allows breakthrough performance for message passing algorithms on arbitrary graphs. We show for the first time that the classical computer vision problem of bundle adjustment (BA) can be solved extremely fast on a graph processor using Gaussian Belief Propagation. Our simple but fully parallel implementation uses the 1216 cores on a single IPU chip to, for instance, solve a real BA problem with 125 keyframes and 1919 points in under 40ms, compared to 1450ms for the Ceres CPU library. Further code optimisation will surely increase this difference on static problems, but we argue that the real promise of graph processing is for flexible in-place optimisation of general, dynamically changing factor graphs representing Spatial AI problems. We give indications of this with experiments showing the ability of GBP to efficiently solve incremental SLAM problems, and deal with robust cost functions and different types of factors.
Even though the Convolutional Neural Networks (CNN) has shown superior results in the field of computer vision, it is still a challenging task to implement computer vision algorithms in real-time at the edge, especially using a low-cost IoT device due to high memory consumption and computation complexities in a CNN. Network compression methodologies such as weight pruning, filter pruning, and quantization are used to overcome the above mentioned problem. Even though filter pruning methodology has shown better performances compared to other techniques, irregularity of the number of filters pruned across different layers of a CNN might not comply with majority of the neural computing hardware architectures. In this paper, a novel greedy approach called cluster pruning has been proposed, which provides a structured way of removing filters in a CNN by considering the importance of filters and the underlying hardware architecture. The proposed methodology is compared with the conventional filter pruning algorithm on Pascal-VOC open dataset, and Head-Counting dataset, which is our own dataset developed to detect and count people entering a room. We benchmark our proposed method on three hardware architectures, namely CPU, GPU, and Intel Movidius Neural Computer Stick (NCS) using the popular SSD-MobileNet and SSD-SqueezeNet neural network architectures used for edge-AI vision applications. Results demonstrate that our method outperforms the conventional filter pruning methodology, using both datasets on above mentioned hardware architectures. Furthermore, a low cost IoT hardware setup consisting of an Intel Movidius-NCS is proposed to deploy an edge-AI application using our proposed pruning methodology.
Running deep neural network (DNN) inference on mobile devices, i.e., mobile inference, has become a growing trend, making inference less dependent on network connections and keeping private data locally. The prior studies on optimizing DNNs for mobile inference typically focus on the metric of average inference latency, thus implicitly assuming that mobile inference exhibits little latency variability. In this note, we conduct a preliminary measurement study on the latency variability of DNNs for mobile inference. We show that the inference latency variability can become quite significant in the presence of CPU resource contention. More interestingly, unlike the common belief that the relative performance superiority of DNNs on one device can carry over to another device and/or another level of resource contention, we highlight that a DNN model with a better latency performance than another model can become outperformed by the other model when resource contention be more severe or running on another device. Thus, when optimizing DNN models for mobile inference, only measuring the average latency may not be adequate; instead, latency variability under various conditions should be accounted for, including but not limited to different devices and different levels of CPU resource contention considered in this note.
Lidar sensors are widely used in various applications, ranging from scientific fields over industrial use to integration in consumer products. With an ever growing number of different driver assistance systems, they have been introduced to automotive series production in recent years and are considered an important building block for the practical realisation of autonomous driving. However, due to the potentially large amount of Lidar points per scan, tailored algorithms are required to identify objects (e.g. pedestrians or vehicles) with high precision in a very short time. In this work, we propose an algorithmic approach for real-time instance segmentation of Lidar sensor data. We show how our method leverages the properties of the Euclidean distance to retain three-dimensional measurement information, while being narrowed down to a two-dimensional representation for fast computation. We further introduce what we call "skip connections", to make our approach robust against over-segmentation and improve assignment in cases of partial occlusion. Through detailed evaluation on public data and comparison with established methods, we show how these aspects enable state-of-the-art performance and runtime on a single CPU core.
The geometry automated theorem proving area distinguishes itself by a large number of specific methods and implementations, different approaches (synthetic, algebraic, semi-synthetic) and different goals and applications (from research in the area of artificial intelligence to applications in education). Apart from the usual measures of efficiency (e.g. CPU time), the possibility of visual and/or readable proofs is also an expected output against which the geometry automated theorem provers (GATP) should be measured. The implementation of a competition between GATP would allow to create a test bench for GATP developers to improve the existing ones and to propose new ones. It would also allow to establish a ranking for GATP that could be used by "clients" (e.g. developers of educational e-learning systems) to choose the best implementation for a given intended use.
Energy efficiency of hardware accelerators of deep neural networks (DNN) can be improved by introducing approximate arithmetic circuits. In order to quantify the error introduced by using these circuits and avoid the expensive hardware prototyping, a software emulator of the DNN accelerator is usually executed on CPU or GPU. However, this emulation is typically two or three orders of magnitude slower than a software DNN implementation running on CPU or GPU and operating with standard floating point arithmetic instructions and common DNN libraries. The reason is that there is no hardware support for approximate arithmetic operations on common CPUs and GPUs and these operations have to be expensively emulated. In order to address this issue, we propose an efficient emulation method for approximate circuits utilized in a given DNN accelerator which is emulated on GPU. All relevant approximate circuits are implemented as look-up tables and accessed through a texture memory mechanism of CUDA capable GPUs. We exploit the fact that the texture memory is optimized for irregular read-only access and in some GPU architectures is even implemented as a dedicated cache. This technique allowed us to reduce the inference time of the emulated DNN accelerator approximately 200 times with respect to an optimized CPU version on complex DNNs such as ResNet. The proposed approach extends the TensorFlow library and is available online at https://github.com/ehw-fit/tf-approximate.
Optimization in machine learning, both theoretical and applied, is presently dominated by first-order gradient methods such as stochastic gradient descent. Second-order optimization methods, that involve second derivatives and/or second order statistics of the data, are far less prevalent despite strong theoretical properties, due to their prohibitive computation, memory and communication costs. In an attempt to bridge this gap between theoretical and practical optimization, we present a scalable implementation of a second-order preconditioned method (concretely, a variant of full-matrix Adagrad), that along with several critical algorithmic and numerical improvements, provides significant convergence and wall-clock time improvements compared to conventional first-order methods on state-of-the-art deep models. Our novel design effectively utilizes the prevalent heterogeneous hardware architecture for training deep models, consisting of a multicore CPU coupled with multiple accelerator units. We demonstrate superior performance compared to state-of-the-art on very large learning tasks such as machine translation with Transformers, language modeling with BERT, click-through rate prediction on Criteo, and image classification on ImageNet with ResNet-50.
In this work we present a new framework for neural networks compression with fine-tuning, which we called Neural Network Compression Framework (NNCF). It leverages recent advances of various network compression methods and implements some of them, such as sparsity, quantization, and binarization. These methods allow getting more hardware-friendly models which can be efficiently run on general-purpose hardware computation units (CPU, GPU) or special Deep Learning accelerators. We show that the developed methods can be successfully applied to a wide range of models to accelerate the inference time while keeping the original accuracy. The framework can be used within the training samples, which are supplied with it, or as a standalone package that can be seamlessly integrated into the existing training code with minimal adaptations. Currently, a PyTorch version of NNCF is available as a part of OpenVINO Training Extensions at https://github.com/openvinotoolkit/nncf.
The research interest in specialized hardware accelerators for deep neural networks (DNN) spikes recently owing to their superior performance and efficiency. However, today's DNN accelerators primarily focus on accelerating specific "kernels" such as convolution and matrix multiplication, which are vital but only part of an end-to-end DNN-enabled application. Meaningful speedups over the entire application often require supporting computations that are, while massively parallel, ill-suited to DNN accelerators. Integrating a general-purpose processor such as a CPU or a GPU incurs significant data movement overhead and leads to resource under-utilization on the DNN accelerators. We propose Simultaneous Multi-mode Architecture (SMA), a novel architecture design and execution model that offers general-purpose programmability on DNN accelerators in order to accelerate end-to-end applications. The key to SMA is the temporal integration of the systolic execution model with the GPU-like SIMD execution model. The SMA exploits the common components shared between the systolic-array accelerator and the GPU, and provides lightweight reconfiguration capability to switch between the two modes in-situ. The SMA achieves up to 63% performance improvement while consuming 23% less energy than the baseline Volta architecture with TensorCore.
Creating a database of 21cm brightness temperature signals from the Epoch of Reionisation (EoR) for an array of reionisation histories is a complex and computationally expensive task, given the range of astrophysical processes involved and the possibly high-dimensional parameter space that is to be probed. We utilise a specific type of neural network, a Progressively Growing Generative Adversarial Network (PGGAN), to produce realistic tomography images of the 21cm brightness temperature during the EoR, covering a continuous three-dimensional parameter space that models varying X-ray emissivity, Lyman band emissivity, and ratio between hard and soft X-rays. The GPU-trained network generates new samples at a resolution of $\sim 3'$ in a second (on a laptop CPU), and the resulting global 21cm signal, power spectrum, and pixel distribution function agree well with those of the training data, taken from the 21SSD catalogue \citep{Semelin2017}. Finally, we showcase how a trained PGGAN can be leveraged for the converse task of inferring parameters from 21cm tomography samples via Approximate Bayesian Computation.
This paper considers a convolutional neural network transformation that reduces computation complexity and thus speedups neural network processing. Usage of convolutional neural networks (CNN) is the standard approach to image recognition despite the fact they can be too computationally demanding, for example for recognition on mobile platforms or in embedded systems. In this paper we propose CNN structure transformation which expresses 2D convolution filters as a linear combination of separable filters. It allows to obtain separated convolutional filters by standard training algorithms. We study the computation efficiency of this structure transformation and suggest fast implementation easily handled by CPU or GPU. We demonstrate that CNNs designed for letter and digit recognition of proposed structure show 15% speedup without accuracy loss in industrial image recognition system. In conclusion, we discuss the question of possible accuracy decrease and the application of proposed transformation to different recognition problems. convolutional neural networks, computational optimization, separable filters, complexity reduction.
Cloud Infrastructure as a Service (IaaS) is vulnerable to malware due to its exposure to external adversaries, making it a lucrative attack vector for malicious actors. A datacenter infected with malware can cause data loss and/or major disruptions to service for its users. This paper analyzes and compares various Convolutional Neural Networks (CNNs) for online detection of malware in cloud IaaS. The detection is performed based on behavioural data using process level performance metrics including cpu usage, memory usage, disk usage etc. We have used the state of the art DenseNets and ResNets in effectively detecting malware in online cloud system. CNN are designed to extract features from data gathered from a live malware running on a real cloud environment. Experiments are performed on OpenStack (a cloud IaaS software) testbed designed to replicate a typical 3-tier web architecture. Comparative analysis is performed for different metrics for different CNN models used in this research.
New data processing pipelines and novel network architectures increasingly drive the success of deep learning. In consequence, the industry considers top-performing architectures as intellectual property and devotes considerable computational resources to discovering such architectures through neural architecture search (NAS). This provides an incentive for adversaries to steal these novel architectures; when used in the cloud, to provide Machine Learning as a Service, the adversaries also have an opportunity to reconstruct the architectures by exploiting a range of hardware side channels. However, it is challenging to reconstruct novel architectures and pipelines without knowing the computational graph (e.g., the layers, branches or skip connections), the architectural parameters (e.g., the number of filters in a convolutional layer) or the specific pre-processing steps (e.g. embeddings). In this paper, we design an algorithm that reconstructs the key components of a novel deep learning system by exploiting a small amount of information leakage from a cache side-channel attack, Flush+Reload. We use Flush+Reload to infer the trace of computations and the timing for each computation. Our algorithm then generates candidate computational graphs from the trace and eliminates incompatible candidates through a parameter estimation process. We implement our algorithm in PyTorch and Tensorflow. We demonstrate experimentally that we can reconstruct MalConv, a novel data pre-processing pipeline for malware detection, and ProxylessNAS- CPU, a novel network architecture for the ImageNet classification optimized to run on CPUs, without knowing the architecture family. In both cases, we achieve 0% error. These results suggest hardware side channels are a practical attack vector against MLaaS, and more efforts should be devoted to understanding their impact on the security of deep learning systems.
The miniaturization of transistors down to 5nm and beyond, plus the increasing complexity of integrated circuits, significantly aggravate short channel effects, and demand analysis and optimization of more design corners and modes. Simulators need to model output variables related to circuit timing, power, noise, etc., which exhibit nonlinear behavior. The existing simulation and sign-off tools, based on a combination of closed-form expressions and lookup tables are either inaccurate or slow, when dealing with circuits with more than billions of transistors. In this work, we present CSM-NN, a scalable simulation framework with optimized neural network structures and processing algorithms. CSM-NN is aimed at optimizing the simulation time by accounting for the latency of the required memory query and computation, given the underlying CPU and GPU parallel processing capabilities. Experimental results show that CSM-NN reduces the simulation time by up to $6\times$ compared to a state-of-the-art current source model based simulator running on a CPU. This speedup improves by up to $15\times$ when running on a GPU. CSM-NN also provides high accuracy levels, with less than $2\%$ error, compared to HSPICE.
Widely popular transformer-based NLP models such as BERT and Turing-NLG have enormous capacity trending to billions of parameters. Current execution methods demand brute-force resources such as HBM devices and high speed interconnectivity for data parallelism. In this paper, we introduce a new relay-style execution technique called L2L (layer-to-layer) where at any given moment, the device memory is primarily populated only with the executing layer(s)'s footprint. The model resides in the DRAM memory attached to either a CPU or an FPGA as an entity we call eager param-server (EPS). To overcome the bandwidth issues of shuttling parameters to and from EPS, the model is executed a layer at a time across many micro-batches instead of the conventional method of minibatches over whole model. L2L is implemented using 16GB V100 devices for BERT-Large running it with a device batch size of up to 256. Our results show 45% reduction in memory and 40% increase in the throughput compared to the state-of-the-art baseline. L2L is also able to fit models up to 50 Billion parameters on a machine with a single 16GB V100 and 512GB CPU memory and without requiring any model partitioning. L2L scales to arbitrary depth allowing researchers to develop on affordable devices which is a big step toward democratizing AI. By running the optimizer in the host EPS, we show a new form of mixed precision for faster throughput and convergence. In addition, the EPS enables dynamic neural architecture approaches by varying layers across iterations. Finally, we also propose and demonstrate a constant memory variation of L2L and we propose future enhancements. This work has been performed on GPUs first, but also targeted towards all high TFLOPS/Watt accelerators.