In this paper, we address the problem of reducing the memory footprint of convolutional network architectures. We introduce a vector quantization method that aims at preserving the quality of the reconstruction of the network outputs rather than its weights. The principle of our approach is that it minimizes the loss reconstruction error for in-domain inputs. Our method only requires a set of unlabelled data at quantization time and allows for efficient inference on CPU by using byte-aligned codebooks to store the compressed weights. We validate our approach by quantizing a high performing ResNet-50 model to a memory size of 5MB (20x compression factor) while preserving a top-1 accuracy of 76.1% on ImageNet object classification and by compressing a Mask R-CNN with a 26x factor.
The proliferation of ubiquitous computing requires energy-efficient as well as secure operation of modern processors. Side channel attacks are becoming a critical threat to security and privacy of devices embedded in modern computing infrastructures. Unintended information leakage via physical signatures such as power consumption, electromagnetic emission (EM) and execution time have emerged as a key security consideration for SoCs. Also, information published on purpose at user privilege level accessible through software interfaces results in software only attacks. In this paper, we used a supervised learning based approach for inferring applications executing on android platform based on features extracted from EM side-channel emissions and software exposed dynamic voltage frequency scaling(DVFS) states. We highlight the importance of machine learning based approach in utilizing these multi-dimensional features on a complex SoC, against profiling-based approaches. We also show that learning the instantaneous frequency states polled from onboard frequency driver (cpufreq) is adequate to identify a known application and flag potentially malicious unknown application. The experimental results on benchmarking applications running on ARMv8 processor in Snapdragon 820 board demonstrates early detection of these apps, and atleast 85% accuracy in detecting unknown applications. Overall, the highlight is to utilize a low-complexity path to application inference attacks through learning instantaneous frequency states pattern of CPU core.
Landmark matching via geodesic shooting is a prerequisite task for numerous registration based applications in biomedicine. Geodesic shooting has been developed as one solution approach and formulates the diffeomorphic registration as an optimal control problem under the Hamiltonian framework. In this framework, with landmark positions q0 fixed, the problem solely depends on the initial momentum p0 and evolves through time steps according to a set of constraint equations. Given an initial p0, the algorithm flows q and p forward through time steps, calculates a loss based on point-set mismatch and kinetic energy, back-propagate through time to calculate gradient on p0 and update it. In the forward and backward pass, a pair-wise kernel on landmark points K and additional intermediate terms have to be calculated and marginalized, leading to O(N2) computational complexity, N being the number of points to be registered. For medical image applications, N maybe in the range of thousands, rendering this operation computationally expensive. In this work we ropose a CUDA implementation based on shared memory reduction. Our implementation achieves nearly 2 orders magnitude speed up compared to a naive CPU-based implementation, in addition to improved numerical accuracy as well as better registration results.
Depth from Small Motion (DfSM) (Ha et al., 2016) is particularly interesting for commercial handheld devices because it allows the possibility to get depth information with minimal user effort and cooperation. Due to speed and memory issue on these devices, the self calibration optimization of the method using Bundle Adjustment (BA) need as little as 10-15 images. Therefore, the optimization tends to take many iterations to converge or may not converge at all in some cases. This work propose a robust initialization for the bundle adjustment using the rank-1 factorization method (Tomasi and Kanade, 1992), (Aguiar and Moura, 1999a). We create a constraint matrix that is rank-1 in a noiseless situation, then use SVD to compute the inverse depth values and the camera motion. We only need about quarter fraction of the bundle adjustment iteration to converge. We also propose grided feature extraction technique so that only important and small features are tracked all over the image frames. This also ensure speedup in the full execution time on the mobile device. For the experiments, we have documented the execution time with the proposed Rank-1 initialization on two mobile device platforms using optimized accelerations with CPU-GPU co-processing. The combination of Rank 1-BA generates more robust depth-map and is significantly faster than using BA alone.
Probabilistic programming languages (PPLs) are receiving widespread attention for performing Bayesian inference in complex generative models. However, applications to science remain limited because of the impracticability of rewriting complex scientific simulators in a PPL, the computational cost of inference, and the lack of scalable implementations. To address these, we present a novel PPL framework that couples directly to existing scientific simulators through a cross-platform probabilistic execution protocol and provides Markov chain Monte Carlo (MCMC) and deep-learning-based inference compilation (IC) engines for tractable inference. To guide IC inference, we perform distributed training of a dynamic 3DCNN--LSTM architecture with a PyTorch-MPI-based framework on 1,024 32-core CPU nodes of the Cori supercomputer with a global minibatch size of 128k: achieving a performance of 450 Tflop/s through enhancements to PyTorch. We demonstrate a Large Hadron Collider (LHC) use-case with the C++ Sherpa simulator and achieve the largest-scale posterior inference in a Turing-complete PPL.
In this work, we introduce a system for real-time generation of drum sounds. This system is composed of two parts: a generative model for drum sounds together with a Max4Live plugin providing intuitive controls on the generative process. The generative model consists of a Conditional Wasserstein autoencoder (CWAE), which learns to generate Mel-scaled magnitude spectrograms of short percussion samples, coupled with a Multi-Head Convolutional Neural Network (MCNN) which estimates the corresponding audio signal from the magnitude spectrogram. The design of this model makes it lightweight, so that it allows one to perform real-time generation of novel drum sounds on an average CPU, removing the need for the users to possess dedicated hardware in order to use this system. We then present our Max4Live interface designed to interact with this generative model. With this setup, the system can be easily integrated into a studio-production environment and enhance the creative process. Finally, we discuss the advantages of our system and how the interaction of music producers with such tools could change the way drum tracks are composed.
In this paper we tackle the problem of recovering the phase of complex linear measurements when only magnitude information is available and we control the input. We are motivated by the recent development of dedicated optics-based hardware for rapid random projections which leverages the propagation of light in random media. A signal of interest $\mathbf{\xi} \in \mathbb{R}^N$ is mixed by a random scattering medium to compute the projection $\mathbf{y} = \mathbf{A} \mathbf{\xi}$, with $\mathbf{A} \in \mathbb{C}^{M \times N}$ being a realization of a standard complex Gaussian iid random matrix. Such optics-based matrix multiplications can be much faster and energy-efficient than their CPU or GPU counterparts, yet two difficulties must be resolved: only the intensity ${|\mathbf{y}|}^2$ can be recorded by the camera, and the transmission matrix $\mathbf{A}$ is unknown. We show that even without knowing $\mathbf{A}$, we can recover the unknown phase of $\mathbf{y}$ for some equivalent transmission matrix with the same distribution as $\mathbf{A}$. Our method is based on two observations: first, conjugating or changing the phase of any row of $\mathbf{A}$ does not change its distribution; and second, since we control the input we can interfere $\mathbf{\xi}$ with arbitrary reference signals. We show how to leverage these observations to cast the measurement phase retrieval problem as a Euclidean distance geometry problem. We demonstrate appealing properties of the proposed algorithm in both numerical simulations and real hardware experiments. Not only does our algorithm accurately recover the missing phase, but it mitigates the effects of quantization and the sensitivity threshold, thus improving the measured magnitudes.
On-device inference of machine learning models for mobile phones is desirable due to its lower latency and increased privacy. Running such a compute-intensive task solely on the mobile CPU, however, can be difficult due to limited computing power, thermal constraints, and energy consumption. App developers and researchers have begun exploiting hardware accelerators to overcome these challenges. Recently, device manufacturers are adding neural processing units into high-end phones for on-device inference, but these account for only a small fraction of hand-held devices. In this paper, we present how we leverage the mobile GPU, a ubiquitous hardware accelerator on virtually every phone, to run inference of deep neural networks in real-time for both Android and iOS devices. By describing our architecture, we also discuss how to design networks that are mobile GPU-friendly. Our state-of-the-art mobile GPU inference engine is integrated into the open-source project TensorFlow Lite and publicly available at https://tensorflow.org/lite.
It has always been a research hotspot to use geographic information to assist the navigation of unmanned aerial vehicles. In this paper, a road-network-based localization method is proposed. We match roads in the measurement images to the reference road vector map, and realize successful localization on areas as large as a whole city. The road network matching problem is treated as a point cloud registration problem under two-dimensional projective transformation, and solved under a hypothesise-and-test framework. To deal with the projective point cloud registration problem, a global projective invariant feature is proposed, which consists of two road intersections augmented with the information of their tangents. We call it two road intersections tuple. We deduce the closed-form solution for determining the alignment transformation from a pair of matching two road intersections tuples. In addition, we propose the necessary conditions for the tuples to match. This can reduce the candidate matching tuples, thus accelerating the search to a great extent. We test all the candidate matching tuples under a hypothesise-and-test framework to search for the best match. The experiments show that our method can localize the target area over an area of 400 within 1 second on a single cpu.
Developing high performance embedded vision applications requires balancing run-time performance with energy constraints. Given the mix of hardware accelerators that exist for embedded computer vision (e.g. multi-core CPUs, GPUs, and FPGAs), and their associated vendor optimized vision libraries, it becomes a challenge for developers to navigate this fragmented solution space. To aid with determining which embedded platform is most suitable for their application, we conduct a comprehensive benchmark of the run-time performance and energy efficiency of a wide range of vision kernels. We discuss rationales for why a given underlying hardware architecture innately performs well or poorly based on the characteristics of a range of vision kernel categories. Specifically, our study is performed for three commonly used HW accelerators for embedded vision applications: ARM57 CPU, Jetson TX2 GPU and ZCU102 FPGA, using their vendor optimized vision libraries: OpenCV, VisionWorks and xfOpenCV. Our results show that the GPU achieves an energy/frame reduction ratio of 1.1-3.2x compared to the others for simple kernels. While for more complicated kernels and complete vision pipelines, the FPGA outperforms the others with energy/frame reduction ratios of 1.2-22.3x. It is also observed that the FPGA performs increasingly better as a vision application's pipeline complexity grows.
This paper introduces VolMap, a real-time approach for the semantic segmentation of a 3D LiDAR surrounding view system in autonomous vehicles. We designed an optimized deep convolution neural network that can accurately segment the point cloud produced by a 360\degree{} LiDAR setup, where the input consists of a volumetric bird-eye view with LiDAR height layers used as input channels. We further investigated the usage of multi-LiDAR setup and its effect on the performance of the semantic segmentation task. Our evaluations are carried out on a large scale 3D object detection benchmark containing a LiDAR cocoon setup, along with KITTI dataset, where the per-point segmentation labels are derived from 3D bounding boxes. We show that VolMap achieved an excellent balance between high accuracy and real-time running on CPU.
Hyperspectral image (HSI) classification has been widely adopted in applications involving remote sensing imagery analysis which require high classification accuracy and real-time processing speed. Methods based on Convolutional neural networks (CNNs) have been proven to achieve state-of-the-art accuracy in classifying HSIs. However, CNN models are often too computationally intensive to achieve real-time response due to the high dimensional nature of HSI, compared to traditional methods such as Support Vector Machines (SVMs). Besides, previous CNN models used in HSI are not specially designed for efficient implementation on embedded devices such as FPGAs. This paper proposes a novel CNN-based algorithm for HSI classification which takes into account hardware efficiency. A customized architecture which enables the proposed algorithm to be mapped effectively onto FPGA resources is then proposed to support real-time on-board classification with low power consumption. Implementation results show that our proposed accelerator on a Xilinx Zynq 706 FPGA board achieves more than 70x faster than an Intel 8-core Xeon CPU and 3x faster than an NVIDIA GeForce 1080 GPU. Compared to previous SVM-based FPGA accelerators, we achieve comparable processing speed but provide a much higher classification accuracy.
Today's cloud service architectures follow a "one size fits all" deployment strategy where the same service version instantiation is provided to the end users. However, consumers are broad and different applications have different accuracy and responsiveness requirements, which as we demonstrate renders the "one size fits all" approach inefficient in practice. We use a production-grade speech recognition engine, which serves several thousands of users, and an open source computer vision based system, to explain our point. To overcome the limitations of the "one size fits all" approach, we recommend Tolerance Tiers where each MLaaS tier exposes an accuracy/responsiveness characteristic, and consumers can programmatically select a tier. We evaluate our proposal on the CPU-based automatic speech recognition (ASR) engine and cutting-edge neural networks for image classification deployed on both CPUs and GPUs. The results show that our proposed approach provides an MLaaS cloud service architecture that can be tuned by the end API user or consumer to outperform the conventional "one size fits all" approach.
Time Series Classification (TSC) has seen enormous progress over the last two decades. HIVE-COTE (Hierarchical Vote Collective of Transformation-based Ensembles) is the current state of the art in terms of classification accuracy. HIVE-COTE recognizes that time series data are a specific data type for which the traditional attribute-value representation, used predominantly in machine learning, fails to provide a relevant representation. HIVE-COTE combines multiple types of classifiers: each extracting information about a specific aspect of a time series, be it in the time domain, frequency domain or summarization of intervals within the series. However, HIVE-COTE (and its predecessor, FLAT-COTE) is often infeasible to run on even modest amounts of data. For instance, training HIVE-COTE on a dataset with only 1,500 time series can require 8 days of CPU time. It has polynomial runtime with respect to the training set size, so this problem compounds as data quantity increases. We propose a novel TSC algorithm, TS-CHIEF (Time Series Combination of Heterogeneous and Integrated Embedding Forest), which rivals HIVE-COTE in accuracy but requires only a fraction of the runtime. TS-CHIEF constructs an ensemble classifier that integrates the most effective embeddings of time series that research has developed in the last decade. It uses tree-structured classifiers to do so efficiently. We assess TS-CHIEF on 85 datasets of the University of California Riverside (UCR) archive, where it achieves state-of-the-art accuracy with scalability and efficiency. We demonstrate that TS-CHIEF can be trained on 130k time series in 2 days, a data quantity that is beyond the reach of any TSC algorithm with comparable accuracy.
CPU branch prediction has hit a wall--existing techniques achieve near-perfect accuracy on 99% of static branches, and yet the mispredictions that remain hide major performance gains. In a companion report, we show that a primary source of mispredictions is a handful of systematically hard-to-predict branches (H2Ps), e.g. just 10 static instructions per SimPoint phase in SPECint 2017. The lost opportunity posed by these mispredictions is significant to the CPU: 14.0% in instructions-per-cycle (IPC) on Intel SkyLake and 37.4% IPC when the pipeline is scaled four-fold, on par with gains from process technology. However, up to 80% of this upside is unreachable by the best known branch predictors, even when afforded exponentially more resources. New approaches are needed, and machine learning (ML) provides a palette of powerful predictors. A growing body of work has shown that ML models are deployable within the microarchitecture to optimize hardware at runtime, and are one way to customize CPUs post-silicon by training to customer applications. We develop this scenario for branch prediction using convolutional neural networks (CNNs) to boost accuracy for H2Ps. Step-by-step, we (1) map CNNs to the global history data used by existing branch predictors; (2) show how CNNs improve H2P prediction in SPEC 2017; (3) adapt 2-bit CNN inference to the constraints of current branch prediction units; and (4) establish that CNN helper predictors are reusable across application executions on different inputs, enabling us to amortize offline training and deploy ML pattern matching to improve IPC.
Speaker embeddings are continuous-value vector representations that allow easy comparison between voices of speakers with simple geometric operations. Among others, i-vector and x-vector have emerged as the mainstream methods for speaker embedding. In this paper, we illustrate the use of modern computation platform to harness the benefit of GPU acceleration for i-vector extraction. In particular, we achieve an acceleration of 3000 times in frame posterior computation compared to real time and 25 times in training the i-vector extractor compared to the CPU baseline from Kaldi toolkit. This significant speed-up allows the exploration of ideas that were hitherto impossible. In particular, we show that it is beneficial to update the universal background model (UBM) and re-compute frame alignments while training the i-vector extractor. Additionally, we are able to study different variations of i-vector extractors more rigorously than before. In this process, we reveal some undocumented details of Kaldi's i-vector extractor and show that it outperforms the standard formulation by a margin of 1 to 2% when tested with VoxCeleb speaker verification protocol. All of our findings are asserted by ensemble averaging the results from multiple runs with random start.
In recent years, Convolutional Neural Networks (CNNs) have shown superior capability in visual learning tasks. While accuracy-wise CNNs provide unprecedented performance, they are also known to be computationally intensive and energy demanding for modern computer systems. In this paper, we propose Virtual Pooling (ViP), a model-level approach to improve speed and energy consumption of CNN-based image classification and object detection tasks, with a provable error bound. We show the efficacy of ViP through experiments on four CNN models, three representative datasets, both desktop and mobile platforms, and two visual learning tasks, i.e., image classification and object detection. For example, ViP delivers 2.1x speedup with less than 1.5% accuracy degradation in ImageNet classification on VGG-16, and 1.8x speedup with 0.025 mAP degradation in PASCAL VOC object detection with Faster-RCNN. ViP also reduces mobile GPU and CPU energy consumption by up to 55% and 70%, respectively. As a complementary method to existing acceleration approaches, ViP achieves 1.9x speedup on ThiNet leading to a combined speedup of 5.23x on VGG-16. Furthermore, ViP provides a knob for machine learning practitioners to generate a set of CNN models with varying trade-offs between system speed/energy consumption and accuracy to better accommodate the requirements of their tasks. Code is available at https://github.com/cmu-enyac/VirtualPooling.
Porting code from CPU to GPU is costly and time-consuming; Unless much time is invested in development and optimization, it is not obvious, a priori, how much speed-up is achievable or how much room is left for improvement. Knowing the potential speed-up a priori can be very useful: It can save hundreds of engineering hours, help programmers with prioritization and algorithm selection. We aim to address this problem using machine learning in a supervised setting, using solely the single-threaded source code of the program, without having to run or profile the code. We propose a static analysis-based cross-architecture performance prediction framework (Static XAPP) which relies solely on program properties collected using static analysis of the CPU source code and predicts whether the potential speed-up is above or below a given threshold. We offer preliminary results that show we can achieve 94% accuracy in binary classification, in average, across different thresholds
As the performance of computer systems stagnates due to the end of Moore's Law, there is a need for new models that can understand and optimize the execution of general purpose code. While there is a growing body of work on using Graph Neural Networks (GNNs) to learn representations of source code, these representations do not understand how code dynamically executes. In this work, we propose a new approach to use GNNs to learn fused representations of general source code and its execution. Our approach defines a multi-task GNN over low-level representations of source code and program state (i.e., assembly code and dynamic memory states), converting complex source code constructs and complex data structures into a simpler, more uniform format. We show that this leads to improved performance over similar methods that do not use execution and it opens the door to applying GNN models to new tasks that would not be feasible from static code alone. As an illustration of this, we apply the new model to challenging dynamic tasks (branch prediction and prefetching) from the SPEC CPU benchmark suite, outperforming the state-of-the-art by 26% and 45% respectively. Moreover, we use the learned fused graph embeddings to demonstrate transfer learning with high performance on an indirectly related task (algorithm classification).
Given an input graph G and a node v in G, homogeneous network embedding (HNE) maps the graph structure in the vicinity of v to a compact, fixed-dimensional feature vector. This paper focuses on HNE for massive graphs, e.g., with billions of edges. On this scale, most existing approaches fail, as they incur either prohibitively high costs, or severely compromised result utility. Our proposed solution, called Node-Reweighted PageRank (NRP), is based on a classic idea of deriving embedding vectors from pairwise personalized PageRank (PPR) values. Our contributions are twofold: first, we design a simple and efficient baseline HNE method based on PPR that is capable of handling billion-edge graphs on commodity hardware; second and more importantly, we identify an inherent drawback of vanilla PPR, and address it in our main proposal NRP. Specifically, PPR was designed for a very different purpose, i.e., ranking nodes in G based on their relative importance from a source node's perspective. In contrast, HNE aims to build node embeddings considering the whole graph. Consequently, node embeddings derived directly from PPR are of suboptimal utility. The proposed NRP approach overcomes the above deficiency through an effective and efficient node reweighting algorithm, which augments PPR values with node degree information, and iteratively adjusts embedding vectors accordingly. Overall, NRP takes O(mlogn) time and O(m) space to compute all node embeddings for a graph with m edges and n nodes. Our extensive experiments that compare NRP against 18 existing solutions over 7 real graphs demonstrate that NRP achieves higher result utility than all the solutions for link prediction, graph reconstruction and node classification, while being up to orders of magnitude faster. In particular, on a billion-edge Twitter graph, NRP terminates within 4 hours, using a single CPU core.