Few-shot classification is a challenge in machine learning where the goal is to train a classifier using a very limited number of labeled examples. This scenario is likely to occur frequently in real life, for example when data acquisition or labeling is expensive. In this work, we consider the problem of post-labeled few-shot unsupervised learning, a classification task where representations are learned in an unsupervised fashion, to be later labeled using very few annotated examples. We argue that this problem is very likely to occur on the edge, when the embedded device directly acquires the data, and the expert needed to perform labeling cannot be prompted often. To address this problem, we consider an algorithm consisting of the concatenation of transfer learning with clustering using Self-Organizing Maps (SOMs). We introduce a TensorFlow-based implementation to speed-up the process in multi-core CPUs and GPUs. Finally, we demonstrate the effectiveness of the method using standard off-the-shelf few-shot classification benchmarks.
One of the main issue in robotics is the lack of embedded computational power. Recently, state of the art algorithms providing a better understanding of the surroundings (Object detection, skeleton tracking, etc.) are requiring more and more computational power. The lack of embedded computational power is more significant in mass-produced robots because of the difficulties to follow the increasing computational requirements of state of the art algorithms. The integration of an additional GPU allows to overcome this lack of embedded computational power. We introduce in this paper a prototype of Pepper with an embedded GPU, but also with an additional 3D camera on the head of the robot and plugged to the late GPU. This prototype, called Adapted Pepper, was built for the European project called MuMMER (MultiModal Mall Entertainment Robot) in order to embed algorithms like OpenPose, YOLO or to process sensors information and, in all cases, avoid network dependency for deported computation.
Automatic segmentation of multi-sequence (multi-modal) cardiac MR (CMR) images plays a significant role in diagnosis and management for a variety of cardiac diseases. However, the performance of relevant algorithms is significantly affected by the proper fusion of the multi-modal information. Furthermore, particular diseases, such as myocardial infarction, display irregular shapes on images and occupy small regions at random locations. These facts make pathology segmentation of multi-modal CMR images a challenging task. In this paper, we present the Max-Fusion U-Net that achieves improved pathology segmentation performance given aligned multi-modal images of LGE, T2-weighted, and bSSFP modalities. Specifically, modality-specific features are extracted by dedicated encoders. Then they are fused with the pixel-wise maximum operator. Together with the corresponding encoding features, these representations are propagated to decoding layers with U-Net skip-connections. Furthermore, a spatial-attention module is applied in the last decoding layer to encourage the network to focus on those small semantically meaningful pathological regions that trigger relatively high responses by the network neurons. We also use a simple image patch extraction strategy to dynamically resample training examples with varying spacial and batch sizes. With limited GPU memory, this strategy reduces the imbalance of classes and forces the model to focus on regions around the interested pathology. It further improves segmentation accuracy and reduces the mis-classification of pathology. We evaluate our methods using the Myocardial pathology segmentation (MyoPS) combining the multi-sequence CMR dataset which involves three modalities. Extensive experiments demonstrate the effectiveness of the proposed model which outperforms the related baselines.
Recent work on Renyi Differential Privacy has shown the feasibility of applying differential privacy to deep learning tasks. Despite their promise, however, differentially private deep networks often lag far behind their non-private counterparts in accuracy, showing the need for more research in model architectures, optimizers, etc. One of the barriers to this expanded research is the training time -- often orders of magnitude larger than training non-private networks. The reason for this slowdown is a crucial privacy-related step called "per-example gradient clipping" whose naive implementation undoes the benefits of batch training with GPUs. By analyzing the back-propagation equations we derive new methods for per-example gradient clipping that are compatible with auto-differentiation (e.g., in PyTorch and TensorFlow) and provide better GPU utilization. Our implementation in PyTorch showed significant training speed-ups (by factors of 54x - 94x for training various models with batch sizes of 128). These techniques work for a variety of architectural choices including convolutional layers, recurrent networks, attention, residual blocks, etc.
We present Qibo, a new open-source software for fast evaluation of quantum circuits and adiabatic evolution which takes full advantage of hardware accelerators. The growing interest in quantum computing and the recent developments of quantum hardware devices motivates the development of new advanced computational tools focused on performance and usage simplicity. In this work we introduce a new quantum simulation framework that enables developers to delegate all complicated aspects of hardware or platform implementation to the library so they can focus on the problem and quantum algorithms at hand. This software is designed from scratch with simulation performance, code simplicity and user friendly interface as target goals. It takes advantage of hardware acceleration such as multi-threading CPU, single GPU and multi-GPU devices.
We consider the question: what is the abstraction that should be implemented by the computational engine of a machine learning system? Current machine learning systems typically push whole tensors through a series of compute kernels such as matrix multiplications or activation functions, where each kernel runs on an AI accelerator (ASIC) such as a GPU. This implementation abstraction provides little built-in support for ML systems to scale past a single machine, or for handling large models with matrices or tensors that do not easily fit into the RAM of an ASIC. In this paper, we present an alternative implementation abstraction called the tensor relational algebra (TRA). The TRA is a set-based algebra based on the relational algebra. Expressions in the TRA operate over binary tensor relations, where keys are multi-dimensional arrays and values are tensors. The TRA is easily executed with high efficiency in a parallel or distributed environment, and amenable to automatic optimization. Our empirical study shows that the optimized TRA-based back-end can significantly outperform alternatives for running ML workflows in distributed clusters.
Network pruning can reduce the high computation cost of deep neural network (DNN) models. However, to maintain their accuracies, sparse models often carry randomly-distributed weights, leading to irregular computations. Consequently, sparse models cannot achieve meaningful speedup on commodity hardware (e.g., GPU) built for dense matrix computations. As such, prior works usually modify or design completely new sparsity-optimized architectures for exploiting sparsity. We propose an algorithm-software co-designed pruning method that achieves latency speedups on existing dense architectures. Our work builds upon the insight that the matrix multiplication generally breaks the large matrix into multiple smaller tiles for parallel execution. We propose a tiling-friendly "tile-wise" sparsity pattern, which maintains a regular pattern at the tile level for efficient execution but allows for irregular, arbitrary pruning at the global scale to maintain the high accuracy. We implement and evaluate the sparsity pattern on GPU tensor core, achieving a 1.95x speedup over the dense model.
The Eulerian fluid simulation is an important HPC application. The neural network has been applied to accelerate it. The current methods that accelerate the fluid simulation with neural networks lack flexibility and generalization. In this paper, we tackle the above limitation and aim to enhance the applicability of neural networks in the Eulerian fluid simulation. We introduce Smartfluidnet, a framework that automates model generation and application. Given an existing neural network as input, Smartfluidnet generates multiple neural networks before the simulation to meet the execution time and simulation quality requirement. During the simulation, Smartfluidnet dynamically switches the neural networks to make the best efforts to reach the user requirement on simulation quality. Evaluating with 20,480 input problems, we show that Smartfluidnet achieves 1.46x and 590x speedup comparing with a state-of-the-art neural network model and the original fluid simulation respectively on an NVIDIA Titan X Pascal GPU, while providing better simulation quality than the state-of-the-art model.
Face super-resolution (SR) has become an indispensable function in security solutions such as video surveillance and identification system, but the distortion in facial components is a great challenge in it. Most state-of-the-art methods have utilized facial priors with deep neural networks. These methods require extra labels, longer training time, and larger computation memory. In this paper, we propose a novel Edge and Identity Preserving Network for Face SR Network, named as EIPNet, to minimize the distortion by utilizing a lightweight edge block and identity information. We present an edge block to extract perceptual edge information, and concatenate it to the original feature maps in multiple scales. This structure progressively provides edge information in reconstruction to aggregate local and global structural information. Moreover, we define an identity loss function to preserve identification of SR images. The identity loss function compares feature distributions between SR images and their ground truth to recover identities in SR images. In addition, we provide a luminance-chrominance error (LCE) to separately infer brightness and color information in SR images. The LCE method not only reduces the dependency of color information by dividing brightness and color components but also enables our network to reflect differences between SR images and their ground truth in two color spaces of RGB and YUV. The proposed method facilitates the proposed SR network to elaborately restore facial components and generate high quality 8x scaled SR images with a lightweight network structure. Furthermore, our network is able to reconstruct an 128x128 SR image with 215 fps on a GTX 1080Ti GPU. Extensive experiments demonstrate that our network qualitatively and quantitatively outperforms state-of-the-art methods on two challenging datasets: CelebA and VGGFace2.
Recent advancements in machine learning algorithms, especially the development of Deep Neural Networks (DNNs) have transformed the landscape of Artificial Intelligence (AI). With every passing day, deep learning based methods are applied to solve new problems with exceptional results. The portal to the real world is the edge. The true impact of AI can only be fully realized if we can have AI agents continuously interacting with the real world and solving everyday problems. Unfortunately, high compute and memory requirements of DNNs acts a huge barrier towards this vision. Today we circumvent this problem by deploying special purpose inference hardware on the edge while procuring trained models from the cloud. This approach, however, relies on constant interaction with the cloud for transmitting all the data, training on massive GPU clusters, and downloading updated models. This is challenging for bandwidth, privacy, and constant connectivity concerns that autonomous agents may exhibit. In this paper we evaluate techniques for enabling adaptive intelligence on edge devices with zero interaction with any high-end cloud/server. We build a prototype distributed system of Raspberry Pis communicating via WiFi running NeuroEvolutionary (NE) learning and inference. We evaluate the performance of such a collaborative system and detail the compute/communication characteristics of different arrangements of the system that trade-off parallelism versus communication. Using insights from our analysis, we also propose algorithmic modifications to reduce communication by up to 3.6x during the learning phase to enhance scalability even further and match performance of higher end computing devices at scale. We believe that these insights will enable algorithm-hardware co-design efforts for enabling continuous learning on the edge.
Writing programs for heterogeneous platforms optimized for high performance is hard since this requires the code to be tuned at a low level with architecture-specific optimizations that are most times based on fundamentally differing programming paradigms and languages. OpenVX promises to solve this issue for computer vision applications with a royalty-free industry standard that is based on a graph-execution model. Yet, the OpenVX' algorithm space is constrained to a small set of vision functions. This hinders accelerating computations that are not included in the standard. In this paper, we analyze OpenVX vision functions to find an orthogonal set of computational abstractions. Based on these abstractions, we couple an existing Domain-Specific Language (DSL) back end to the OpenVX environment and provide language constructs to the programmer for the definition of user-defined nodes. In this way, we enable optimizations that are not possible to detect with OpenVX graph implementations using the standard computer vision functions. These optimizations can double the throughput on an Nvidia GTX GPU and decrease the resource usage of a Xilinx Zynq FPGA by 50% for our benchmarks. Finally, we show that our proposed compiler framework, called HipaccVX, can achieve better results than the state-of-the-art approaches Nvidia VisionWorks and Halide-HLS.
Low light images suffer from severe noise, low brightness, low contrast, etc. In previous researches, many image enhancement methods have been proposed, but few methods can deal with these problems simultaneously. In this paper, to solve these problems simultaneously, we propose a low light image enhancement method that can combined with supervised learning and previous HSV (Hue, Saturation, Value) or Retinex model based image enhancement methods. First, we analyse the relationship between the HSV color space and the Retinex theory, and show that the V channel (V channel in HSV color space, equals the maximum channel in RGB color space) of the enhanced image can well represent the contrast and brightness enhancement process. Then, a data-driven conditional re-enhancement network (denoted as CRENet) is proposed. The network takes low light images as input and the enhanced V channel as condition, then it can re-enhance the contrast and brightness of the low light image and at the same time reduce noise and color distortion. It should be noted that during the training process, any paired images with different exposure time can be used for training, and there is no need to carefully select the supervised images which will save a lot. In addition, it takes less than 20 ms to process a color image with the resolution 400*600 on a 2080Ti GPU. Finally, some comparative experiments are implemented to prove the effectiveness of the method. The results show that the method proposed in this paper can significantly improve the quality of the enhanced image, and by combining with other image contrast enhancement methods, the final enhancement result can even be better than the reference image in contrast and brightness. (Code will be available at https://github.com/hitzhangyu/image-enhancement-with-denoise)
Graph neural networks (GNNs) are gaining increasing popularity as a promising approach to machine learning on graphs. Unlike traditional graph workloads where each vertex/edge is associated with a scalar, GNNs attach a feature tensor to each vertex/edge. This additional feature dimension, along with consequently more complex vertex- and edge-wise computations, has enormous implications on locality and parallelism, which existing graph processing systems fail to exploit. This paper proposes FeatGraph to accelerate GNN workloads by co-optimizing graph traversal and feature dimension computation. FeatGraph provides a flexible programming interface to express diverse GNN models by composing coarse-grained sparse templates with fine-grained user-defined functions (UDFs) on each vertex/edge. FeatGraph incorporates optimizations for graph traversal into the sparse templates and allows users to specify optimizations for UDFs with a feature dimension schedule (FDS). FeatGraph speeds up end-to-end GNN training and inference by up to 32x on CPU and 7x on GPU.
With the advances of fifth-generation (5G) cellular networks technology, many studies and work have been carried out on how to build 5G networks for smart cities. In the previous research, street lighting poles and smart light poles are capable of being a 5G access point. In order to determine the position of the points, this paper discusses a new way to identify poles based on Mask R-CNN, which extends Fast R-CNNs by making it employ recursive Bayesian filtering and perform proposal propagation and reuse. The dataset contains 3,000 high-resolution images from google map. To make training faster, we used a very efficient GPU implementation of the convolution operation. We achieved a train error rate of 7.86% and a test error rate of 32.03%. At last, we used the immune algorithm to set 5G poles in the smart cities.
Dynamic resource management has become one of the major areas of research in modern computer and communication system design due to lower power consumption and higher performance demands. The number of integrated cores, level of heterogeneity and amount of control knobs increase steadily. As a result, the system complexity is increasing faster than our ability to optimize and dynamically manage the resources. Moreover, offline approaches are sub-optimal due to workload variations and large volume of new applications unknown at design time. This paper first reviews recent online learning techniques for predicting system performance, power, and temperature. Then, we describe the use of predictive models for online control using two modern approaches: imitation learning (IL) and an explicit nonlinear model predictive control (NMPC). Evaluations on a commercial mobile platform with 16 benchmarks show that the IL approach successfully adapts the control policy to unknown applications. The explicit NMPC provides 25% energy savings compared to a state-of-the-art algorithm for multi-variable power management of modern GPU sub-systems.
The quality of non-default ranking on e-commerce platforms, such as based on ascending item price or descending historical sales volume, often suffers from acute relevance problems, since the irrelevant items are much easier to be exposed at the top of the ranking results. In this work, we propose a two-stage ranking scheme, which first recalls wide range of candidate items through refined query/title keyword matching, and then classifies the recalled items using BERT-Large fine-tuned on human label data. We also implemented parallel prediction on multiple GPU hosts and a C++ tokenization custom op of Tensorflow. In this data challenge, our model won the 1st place in the supervised phase (based on overall F1 score) and 2nd place in the final phase (based on average per query F1 score).
We present a novel multi-view training framework and CNN architecture for combining information from multiple overlapping satellite images and noisy training labels derived from OpenStreetMap (OSM) to semantically label buildings and roads across large geographic regions (100 km$^2$). Our approach to multi-view semantic segmentation yields a 4-7% improvement in the per-class IoU scores compared to the traditional approaches that use the views independently of one another. A unique (and, perhaps, surprising) property of our system is that modifications that are added to the tail-end of the CNN for learning from the multi-view data can be discarded at the time of inference with a relatively small penalty in the overall performance. This implies that the benefits of training using multiple views are absorbed by all the layers of the network. Additionally, our approach only adds a small overhead in terms of the GPU-memory consumption even when training with as many as 32 views per scene. The system we present is end-to-end automated, which facilitates comparing the classifiers trained directly on true orthophotos vis-a-vis first training them on the off-nadir images and subsequently translating the predicted labels to geographical coordinates. With no human supervision, our IoU scores for the buildings and roads classes are 0.8 and 0.64 respectively which are better than state-of-the-art approaches that use OSM labels and that are not completely automated.
Despite the widespread adoption of deep learning for machine translation, it is still expensive to develop high-quality translation models. In this work, we investigate the use of pre-trained models, such as T5 for Portuguese-English and English-Portuguese translation tasks using low-cost hardware. We explore the use of Portuguese and English pre-trained language models and propose an adaptation of the English tokenizer to represent Portuguese characters, such as diaeresis, acute and grave accents. We compare our models to the Google Translate API and MarianMT on a subset of the ParaCrawl dataset, as well as to the winning submission to the WMT19 Biomedical Translation Shared Task. We also describe our submission to the WMT20 Biomedical Translation Shared Task. Our results show that our models have a competitive performance to state-of-the-art models while being trained on modest hardware (a single 8GB gaming GPU for nine days). Our data, models and code are available at https://github.com/unicamp-dl/Lite-T5-Translation.
X-ray diffraction based microscopy techniques such as High Energy Diffraction Microscopy rely on knowledge of the position of diffraction peaks with high precision. These positions are typically computed by fitting the observed intensities in area detector data to a theoretical peak shape such as pseudo-Voigt. As experiments become more complex and detector technologies evolve, the computational cost of such peak detection and shape fitting becomes the biggest hurdle to the rapid analysis required for real-time feedback during in-situ experiments. To this end, we propose BraggNN, a deep learning-based method that can determine peak positions much more rapidly than conventional pseudo-Voigt peak fitting. When applied to a test dataset, BraggNN gives errors of less than 0.29 and 0.57 pixels, relative to the conventional method, for 75% and 95% of the peaks, respectively. When applied to a real experimental dataset, a 3D reconstruction that used peak positions computed by BraggNN yields 15% better results on average as compared to a reconstruction obtained using peak positions determined using conventional 2D pseudo-Voigt fitting. Recent advances in deep learning method implementations and special-purpose model inference accelerators allow BraggNN to deliver enormous performance improvements relative to the conventional method, running, for example, more than 200 times faster than a conventional method on a consumer-class GPU card with out-of-the-box software.
Hardware-aware neural architecture designs have been predominantly focusing on optimizing model performance on single hardware and model development complexity, where another important factor, model deployment complexity, has been largely ignored. In this paper, we argue that, for applications that may be deployed on multiple hardware, having different single-hardware models across the deployed hardware makes it hard to guarantee consistent outputs across hardware and duplicates engineering work for debugging and fixing. To minimize such deployment cost, we propose an alternative solution, multi-hardware models, where a single architecture is developed for multiple hardware. With thoughtful search space design and incorporating the proposed multi-hardware metrics in neural architecture search, we discover multi-hardware models that give state-of-the-art (SoTA) performance across multiple hardware in both average and worse case scenarios. For performance on individual hardware, the single multi-hardware model yields similar or better results than SoTA performance on accelerators like GPU, DSP and EdgeTPU which was achieved by different models, while having similar performance with MobilenetV3 Large Minimalistic model on mobile CPU.