Quantization has emerged to be an effective way to significantly boost the performance of deep neural networks (DNNs) by utilizing low-bit computations. Despite having lower numerical precision, quantized DNNs are able to reduce both memory bandwidth and computation cycles with little losses of accuracy. Integer GEMM (General Matrix Multiplication) is critical to running quantized DNN models efficiently, as GEMM operations often dominate the computations in these models. Various approaches have been developed by leveraging techniques such as vectorization and memory layout to improve the performance of integer GEMM. However, these existing approaches are not fast enough in certain scenarios. We developed NGEMM, a compiler-based GEMM implementation for accelerating lower-precision training and inference. NGEMM has better use of the vector units by avoiding unnecessary vector computation that is introduced during tree reduction. We compared NGEMM's performance with the state-of-art BLAS libraries such as MKL. Our experimental results showed that NGEMM outperformed MKL non-pack and pack version by an average of 1.86x and 1.16x, respectively. We have applied NGEMM to a number of production services in Microsoft.
Deep learning algorithms have been applied very successfully in recent years to a range of problems out of reach for classical solution paradigms. Nevertheless, there is no completely rigorous mathematical error and convergence analysis which explains the success of deep learning algorithms. The error of a deep learning algorithm can in many situations be decomposed into three parts, the approximation error, the generalization error, and the optimization error. In this work we estimate for a certain deep learning algorithm each of these three errors and combine these three error estimates to obtain an overall error analysis for the deep learning algorithm under consideration. In particular, we thereby establish convergence with a suitable convergence speed for the overall error of the deep learning algorithm under consideration. Our convergence speed analysis is far from optimal and the convergence speed that we establish is rather slow, increases exponentially in the dimensions, and, in particular, suffers from the curse of dimensionality. The main contribution of this work is, instead, to provide a full error analysis (i) which covers each of the three different sources of errors usually emerging in deep learning algorithms and (ii) which merges these three sources of errors into one overall error estimate for the considered deep learning algorithm.
This work proposes the use of Bayesian approximations of uncertainty from deep learning in a robot planner, showing that this produces more cautious actions in safety-critical scenarios. The case study investigated is motivated by a setup where an aerial robot acts as a "scout" for a ground robot. This is useful when the below area is unknown or dangerous, with applications in space exploration, military, or search-and-rescue. Images taken from the aerial view are used to provide a less obstructed map to guide the navigation of the robot on the ground. Experiments are conducted using a deep learning semantic image segmentation, followed by a path planner based on the resulting cost map, to provide an empirical analysis of the proposed method. A comparison with similar approaches is presented to portray the usefulness of certain techniques, or variations within a technique, in similar experimental settings. The method is analyzed to assess the impact of variations in the uncertainty extraction, as well as the absence of an uncertainty metric, on the overall system with the use of a defined metric which measures surprise to the planner. The analysis is performed on multiple datasets, showing a similar trend of lower surprise when uncertainty information is incorporated in the planning, given threshold values of the hyperparameters in the uncertainty extraction have been met. We find that taking uncertainty into account leads to paths that could be 18% less risky on an average.
Deep Learning has established itself to be a common occurrence in the business lexicon. The unprecedented success of deep learning in recent years can be attributed to: abundance of data, availability of gargantuan compute capabilities offered by GPUs, and adoption of open-source philosophy by the researchers and industry. Deep neural networks can be decomposed into a series of different operators. MIOpen, AMD's open-source deep learning primitives library for GPUs, provides highly optimized implementations of such operators, shielding researchers from internal implementation details and hence, accelerating the time to discovery. This paper introduces MIOpen and provides details about the internal workings of the library and supported features. MIOpen innovates on several fronts, such as implementing fusion to optimize for memory bandwidth and GPU launch overheads, providing an auto-tuning infrastructure to overcome the large design space of problem configurations, and implementing different algorithms to optimize convolutions for different filter and input sizes. MIOpen is one of the first libraries to publicly support the bfloat16 data-type for convolutions, allowing efficient training at lower precision without the loss of accuracy.
Machine learning as a service (MLaaS), and algorithm marketplaces are on a rise. Data holders can easily train complex models on their data using third party provided learning codes. Training accurate ML models requires massive labeled data and advanced learning algorithms. The resulting models are considered as intellectual property of the model owners and their copyright should be protected. Also, MLaaS needs to be trusted not to embed secret information about the training data into the model, such that it could be later retrieved when the model is deployed. In this paper, we present \emph{membership encoding} for training deep neural networks and encoding the membership information, i.e. whether a data point is used for training, for a subset of training data. Membership encoding has several applications in different scenarios, including robust watermarking for model copyright protection, and also the risk analysis of stealthy data embedding privacy attacks. Our encoding algorithm can determine the membership of significantly redacted data points, and is also robust to model compression and fine-tuning. It also enables encoding a significant fraction of the training set, with negligible drop in the model's prediction accuracy.
Although deep reinforcement learning agents have produced impressive results in many domains, their decision making is difficult to explain to humans. To address this problem, past work has mainly focused on explaining why an action was chosen in a given state. A different type of explanation that is useful is a counterfactual, which deals with "what if?" scenarios. In this work, we introduce the concept of a counterfactual state to help humans gain a better understanding of what would need to change (minimally) in an Atari game image for the agent to choose a different action. We introduce a novel method to create counterfactual states from a generative deep learning architecture. In addition, we evaluate the effectiveness of counterfactual states on human participants who are not machine learning experts. Our user study results suggest that our generated counterfactual states are useful in helping non-expert participants gain a better understanding of an agent's decision making process.
Human Activity Recognition (HAR) using wearable and mobile sensors has gained momentum in last few years, in various fields, such as, healthcare, surveillance, education, entertainment. Nowadays, Edge Computing has emerged to reduce communication latency and network traffic.Edge devices are resource constrained devices and cannot support high computation. In literature, various models have been developed for HAR. In recent years, deep learning algorithms have shown high performance in HAR, but these algorithms require lot of computation making them inefficient to be deployed on edge devices. This paper, proposes a Lightweight Deep Learning Model for HAR requiring less computational power, making it suitable to be deployed on edge devices. The performance of proposed model is tested on the participants six daily activities data. Results show that the proposed model outperforms many of the existing machine learning and deep learning techniques.
This work considers noise removal from images, focusing on the well known K-SVD denoising algorithm. This sparsity-based method was proposed in 2006, and for a short while it was considered as state-of-the-art. However, over the years it has been surpassed by other methods, including the recent deep-learning-based newcomers. The question we address in this paper is whether K-SVD was brought to its peak in its original conception, or whether it can be made competitive again. The approach we take in answering this question is to redesign the algorithm to operate in a supervised manner. More specifically, we propose an end-to-end deep architecture with the exact K-SVD computational path, and train it for optimized denoising. Our work shows how to overcome difficulties arising in turning the K-SVD scheme into a differentiable, and thus learnable, machine. With a small number of parameters to learn and while preserving the original K-SVD essence, the proposed architecture is shown to outperform the classical K-SVD algorithm substantially, and getting closer to recent state-of-the-art learning-based denoising methods. Adopting a broader context, this work touches on themes around the design of deep-learning solutions for image processing tasks, while paving a bridge between classic methods and novel deep-learning-based ones.
Metagenomic studies have increasingly utilized sequencing technologies in order to analyze DNA fragments found in environmental samples.One important step in this analysis is the taxonomic classification of the DNA fragments. Conventional read classification methods require large databases and vast amounts of memory to run, with recent deep learning methods suffering from very large model sizes. We therefore aim to develop a more memory-efficient technique for taxonomic classification. A task of particular interest is abundance estimation in metagenomic samples. Current attempts rely on classifying single DNA reads independently from each other and are therefore agnostic to co-occurence patterns between taxa. In this work, we also attempt to take these patterns into account. We develop a novel memory-efficient read classification technique, combining deep learning and locality-sensitive hashing. We show that this approach outperforms conventional mapping-based and other deep learning methods for single-read taxonomic classification when restricting all methods to a fixed memory footprint. Moreover, we formulate the task of abundance estimation as a Multiple Instance Learning (MIL) problem and we extend current deep learning architectures with two different types of permutation-invariant MIL pooling layers: a) deepsets and b) attention-based pooling. We illustrate that our architectures can exploit the co-occurrence of species in metagenomic read sets and outperform the single-read architectures in predicting the distribution over taxa at higher taxonomic ranks.
Recent advances in computer vision have made accurate, fast and robust measurement of animal behavior a reality. In the past years powerful tools specifically designed to aid the measurement of behavior have come to fruition. Here we discuss how capturing the postures of animals - pose estimation - has been rapidly advancing with new deep learning methods. While challenges still remain, we envision that the fast-paced development of new deep learning tools will rapidly change the landscape of realizable real-world neuroscience.
For an explanation of a deep learning model to be effective, it must provide both insight into a model and suggest a corresponding action in order to achieve some objective. Too often, the litany of proposed explainable deep learning methods stop at the first step, providing practitioners with insight into a model, but no way to act on it. In this paper, we propose contextual decomposition explanation penalization (CDEP), a method which enables practitioners to leverage existing explanation methods in order to increase the predictive accuracy of deep learning models. In particular, when shown that a model has incorrectly assigned importance to some features, CDEP enables practitioners to correct these errors by directly regularizing the provided explanations. Using explanations provided by contextual decomposition (CD) (Murdoch et al., 2018), we demonstrate the ability of our method to increase performance on an array of toy and real datasets.
In this paper we describe a video surveillance system able to detect traffic events in videos acquired by fixed videocameras on highways. The events of interest consist in a specific sequence of situations that occur in the video, as for instance a vehicle stopping on the emergency lane. Hence, the detection of these events requires to analyze a temporal sequence in the video stream. We compare different approaches that exploit architectures based on Recurrent Neural Networks (RNNs) and Convolutional Neural Networks (CNNs). A first approach extracts vectors of features, mostly related to motion, from each video frame and exploits a RNN fed with the resulting sequence of vectors. The other approaches are based directly on the sequence of frames, that are eventually enriched with pixel-wise motion information. The obtained stream is processed by an architecture that stacks a CNN and a RNN, and we also investigate a transfer-learning-based model. The results are very promising and the best architecture will be tested online in real operative conditions.
Financial markets have a vital role in the development of modern society. They allow the deployment of economic resources. Changes in stock prices reflect changes in the market. In this study, we focus on predicting stock prices by deep learning model. This is a challenge task, because there is much noise and uncertainty in information that is related to stock prices. So this work uses sparse autoencoders with one-dimension (1-D) residual convolutional networks which is a deep learning model, to de-noise the data. Long-short term memory (LSTM) is then used to predict the stock price. The prices, indices and macroeconomic variables in past are the features used to predict the next day's price. Experiment results show that 1-D residual convolutional networks can de-noise data and extract deep features better than a model that combines wavelet transforms (WT) and stacked autoencoders (SAEs). In addition, we compare the performances of model with two different forecast targets of stock price: absolute stock price and price rate of change. The results show that predicting stock price through price rate of change is better than predicting absolute prices directly.
Automatic detection of leukemic B-lymphoblast cancer in microscopic images is very challenging due to the complicated nature of histopathological structures. To tackle this issue, an automatic and robust diagnostic system is required for early detection and treatment. In this paper, an automated deep learning-based method is proposed to distinguish between immature leukemic blasts and normal cells. The proposed deep learning based hybrid method, which is enriched by different data augmentation techniques, is able to extract high-level features from input images. Results demonstrate that the proposed model yields better prediction than individual models for Leukemic B-lymphoblast classification with 96.17% overall accuracy, 95.17% sensitivity and 98.58% specificity. Fusing the features extracted from intermediate layers, our approach has the potential to improve the overall classification performance.
Cone-beam breast computed tomography (CT) provides true 3D breast images with isotropic resolution and high-contrast information, detecting calcifications as small as a few hundred microns and revealing subtle tissue differences. However, breast is highly sensitive to x-ray radiation. It is critically important for healthcare to reduce radiation dose. Few-view cone-beam CT only uses a fraction of x-ray projection data acquired by standard cone-beam breast CT, enabling significant reduction of the radiation dose. However, insufficient sampling data would cause severe streak artifacts in CT images reconstructed using conventional methods. In this study, we propose a deep-learning-based method to establish a residual neural network model for the image reconstruction, which is applied for few-view breast CT to produce high quality breast CT images. We respectively evaluate the deep-learning-based image reconstruction using one third and one quarter of x-ray projection views of the standard cone-beam breast CT. Based on clinical breast imaging dataset, we perform a supervised learning to train the neural network from few-view CT images to corresponding full-view CT images. Experimental results show that the deep learning-based image reconstruction method allows few-view breast CT to achieve a radiation dose <6 mGy per cone-beam CT scan, which is a threshold set by FDA for mammographic screening.
Deep learning, through the use of neural networks, has demonstrated remarkable ability to automate many routine tasks when presented with sufficient data for training. The neural network architecture (e.g. number of layers, types of layers, connections between layers, etc.) plays a critical role in determining what, if anything, the neural network is able to learn from the training data. The trend for neural network architectures, especially those trained on ImageNet, has been to grow ever deeper and more complex. The result has been ever increasing accuracy on benchmark datasets with the cost of increased computational demands. In this paper we demonstrate that neural network architectures can be automatically generated, tailored for a specific application, with dual objectives: accuracy of prediction and speed of prediction. Using MENNDL--an HPC-enabled software stack for neural architecture search--we generate a neural network with comparable accuracy to state-of-the-art networks on a cancer pathology dataset that is also $16\times$ faster at inference. The speedup in inference is necessary because of the volume and velocity of cancer pathology data; specifically, the previous state-of-the-art networks are too slow for individual researchers without access to HPC systems to keep pace with the rate of data generation. Our new model enables researchers with modest computational resources to analyze newly generated data faster than it is collected.
In this work, we introduce a deep learning-based polar code construction algorithm. The core idea is to represent the information/frozen bit indices of a polar code as a binary vector which can be interpreted as trainable weights of a neural network (NN). For this, we demonstrate how this binary vector can be relaxed to a soft-valued vector, facilitating the learning process through gradient descent and enabling an efficient code construction. We further show how different polar code design constraints (e.g., code rate) can be taken into account by means of careful binary-to-soft and soft-to-binary conversions, along with rate-adjustment after each learning iteration. Besides its conceptual simplicity, this approach benefits from having the "decoder-in-the-loop", i.e., the nature of the decoder is inherently taken into consideration while learning (designing) the polar code. We show results for belief propagation (BP) decoding over both AWGN and Rayleigh fading channels with considerable performance gains over state-of-the-art construction schemes.
The lack of well-structured metadata annotations complicates there-usability and interpretation of the growing amount of publicly available RNA expression data. The machine learning-based prediction of metadata(data augmentation) can considerably improve the quality of expression data annotation. In this study,we systematically benchmark deep learning (DL) and random forest (RF)-based metadata augmentation of tissue, age, and sex using small RNA (sRNA) expression profiles. We use 4243 annotated sRNA-Seq samples from the small RNA expression atlas (SEA) database to train and test the augmentation performance. In general, the DL machine learner outperforms the RF method in almost all tested cases. The average cross-validated prediction accuracy of the DL algorithm for tissues is 96.5%, for sex is 77%, and for age is 77.2%. The average tissue prediction accuracy for a completely new dataset is 83.1% (DL) and 80.8% (RF). To understand which sRNAs influence DL predictions, we employ backpropagation-based feature importance scores using the DeepLIFT method, which enable us to obtain information on biological relevance of sRNAs.
The lack of well-structured annotations in a growing amount of RNA expression data complicates data interoperability and reusability. Commonly - used text mining methods extract annotations from existing unstructured data descriptions and often provide inaccurate output that requires manual curation. Automatic data-based augmentation (generation of annotations on the base of expression data) can considerably improve the annotation quality and has not been well-studied. We formulate an automatic augmentation of small RNA-seq expression data as a classification problem and investigate deep learning (DL) and random forest (RF) approaches to solve it. We generate tissue and sex annotations from small RNA-seq expression data for tissues and cell lines of homo sapiens. We validate our approach on 4243 annotated small RNA-seq samples from the Small RNA Expression Atlas (SEA) database. The average prediction accuracy for tissue groups is 98% (DL), for tissues - 96.5% (DL), and for sex - 77% (DL). The "one dataset out" average accuracy for tissue group prediction is 83% (DL) and 59% (RF). On average, DL provides better results as compared to RF, and considerably improves classification performance for 'unseen' datasets.
Synthetic data is an increasingly popular tool for training deep learning models, especially in computer vision but also in other areas. In this work, we attempt to provide a comprehensive survey of the various directions in the development and application of synthetic data. First, we discuss synthetic datasets for basic computer vision problems, both low-level (e.g., optical flow estimation) and high-level (e.g., semantic segmentation), synthetic environments and datasets for outdoor and urban scenes (autonomous driving), indoor scenes (indoor navigation), aerial navigation, simulation environments for robotics, applications of synthetic data outside computer vision (in neural programming, bioinformatics, NLP, and more); we also survey the work on improving synthetic data development and alternative ways to produce it such as GANs. Second, we discuss in detail the synthetic-to-real domain adaptation problem that inevitably arises in applications of synthetic data, including synthetic-to-real refinement with GAN-based models and domain adaptation at the feature/model level without explicit data transformations. Third, we turn to privacy-related applications of synthetic data and review the work on generating synthetic datasets with differential privacy guarantees. We conclude by highlighting the most promising directions for further work in synthetic data studies.