While machine learning is traditionally a resource intensive task, embedded systems, autonomous navigation, and the vision of the Internet of Things fuel the interest in resource-efficient approaches. These approaches aim for a carefully chosen trade-off between performance and resource consumption in terms of computation and energy. The development of such approaches is among the major challenges in current machine learning research and key to ensure a smooth transition of machine learning technology from a scientific environment with virtually unlimited computing resources into everyday's applications. In this article, we provide an overview of the current state of the art of machine learning techniques facilitating these real-world requirements. In particular, we focus on resource-efficient inference based on deep neural networks (DNNs), the predominant machine learning models of the past decade. We give a comprehensive overview of the vast literature that can be mainly split into three non-mutually exclusive categories: (i) quantized neural networks, (ii) network pruning, and (iii) structural efficiency. These techniques can be applied during training or as post-processing, and they are widely used to reduce the computational demands in terms of memory footprint, inference speed, and energy efficiency. We also briefly discuss different concepts of embedded hardware for DNNs and their compatibility with machine learning techniques as well as potential for energy and latency reduction. We substantiate our discussion with experiments on well-known benchmark data sets using compression techniques (quantization, pruning) for a set of resource-constrained embedded systems, such as CPUs, GPUs and FPGAs. The obtained results highlight the difficulty of finding good trade-offs between resource efficiency and prediction quality.

This paper introduces neural architecture search (NAS) for the automatic discovery of end-to-end keyword spotting (KWS) models for limited resource environments. We employ a differentiable NAS approach to optimize the structure of convolutional neural networks (CNNs) operating on raw audio waveforms. After a suitable KWS model is found with NAS, we conduct quantization of weights and activations to reduce the memory footprint. We conduct extensive experiments on the Google speech commands dataset. In particular, we compare our end-to-end models to mel-frequency cepstral coefficient (MFCC) based CNNs. For quantization, we compare fixed bitwidth quantization and trained bit-width quantization. Using NAS only, we were able to obtain a highly efficient model with an accuracy of 95.55% using 75.7k parameters and 13.6M operations. Using trained bit-width quantization, the same model achieves a test accuracy of 93.76% while using on average only 2.91 bits per activation and 2.51 bits per weight.

Radar sensors are crucial for environment perception of driver assistance systems as well as autonomous vehicles. With a rising number of radar sensors and the so far unregulated automotive radar frequency band, mutual interference is inevitable and must be dealt with. Algorithms and models operating on radar data are required to run the early processing steps on specialized radar sensor hardware. This specialized hardware typically has strict resource-constraints, i.e. a low memory capacity and low computational power. Convolutional Neural Network (CNN)-based approaches for denoising and interference mitigation yield promising results for radar processing in terms of performance. Regarding resource-constraints, however, CNNs typically exceed the hardware's capacities by far. In this paper we investigate quantization techniques for CNN-based denoising and interference mitigation of radar signals. We analyze the quantization of (i) weights and (ii) activations of different CNN-based model architectures. This quantization results in reduced memory requirements for model storage and during inference. We compare models with fixed and learned bit-widths and contrast two different methodologies for training quantized CNNs, i.e. the straight-through gradient estimator and training distributions over discrete weights. We illustrate the importance of structurally small real-valued base models for quantization and show that learned bit-widths yield the smallest models. We achieve a memory reduction of around 80% compared to the real-valued baseline. Due to practical reasons, however, we recommend the use of 8 bits for weights and activations, which results in models that require only 0.2 megabytes of memory.

We present two methods to reduce the complexity of Bayesian network (BN) classifiers. First, we introduce quantization-aware training using the straight-through gradient estimator to quantize the parameters of BNs to few bits. Second, we extend a recently proposed differentiable tree-augmented naive Bayes (TAN) structure learning approach by also considering the model size. Both methods are motivated by recent developments in the deep learning community, and they provide effective means to trade off between model size and prediction accuracy, which is demonstrated in extensive experiments. Furthermore, we contrast quantized BN classifiers with quantized deep neural networks (DNNs) for small-scale scenarios which have hardly been investigated in the literature. We show Pareto optimal models with respect to model size, number of operations, and test error and find that both model classes are viable options.

This paper introduces neural architecture search (NAS) for the automatic discovery of small models for keyword spotting (KWS) in limited resource environments. We employ a differentiable NAS approach to optimize the structure of convolutional neural networks (CNNs) to maximize the classification accuracy while minimizing the number of operations per inference. Using NAS only, we were able to obtain a highly efficient model with 95.4% accuracy on the Google speech commands dataset with 494.8 kB of memory usage and 19.6 million operations. Additionally, weight quantization is used to reduce the memory consumption even further. We show that weight quantization to low bit-widths (e.g. 1 bit) can be used without substantial loss in accuracy. By increasing the number of input features from 10 MFCC to 20 MFCC we were able to increase the accuracy to 96.3% at 340.1 kB of memory usage and 27.1 million operations.

As a result of the growing size of Deep Neural Networks (DNNs), the gap to hardware capabilities in terms of memory and compute increases. To effectively compress DNNs, quantization and connection pruning are usually considered. However, unconstrained pruning usually leads to unstructured parallelism, which maps poorly to massively parallel processors, and substantially reduces the efficiency of general-purpose processors. Similar applies to quantization, which often requires dedicated hardware. We propose Parameterized Structured Pruning (PSP), a novel method to dynamically learn the shape of DNNs through structured sparsity. PSP parameterizes structures (e.g. channel- or layer-wise) in a weight tensor and leverages weight decay to learn a clear distinction between important and unimportant structures. As a result, PSP maintains prediction performance, creates a substantial amount of sparsity that is structured and, thus, easy and efficient to map to a variety of massively parallel processors, which are mandatory for utmost compute power and energy efficiency. PSP is experimentally validated on the popular CIFAR10/100 and ILSVRC2012 datasets using ResNet and DenseNet architectures, respectively.

While machine learning techniques are traditionally resource intensive, we are currently witnessing an increased interest in hardware and energy efficient approaches. This need for resource-efficient machine learning is primarily driven by the demand for embedded systems and their usage in ubiquitous computing and IoT applications. In this article, we provide a resource-efficient approach for multi-channel speech enhancement based on Deep Neural Networks (DNNs). In particular, we use reduced-precision DNNs for estimating a speech mask from noisy, multi-channel microphone observations. This speech mask is used to obtain either the Minimum Variance Distortionless Response (MVDR) or Generalized Eigenvalue (GEV) beamformer. In the extreme case of binary weights and reduced precision activations, a significant reduction of execution time and memory footprint is possible while still obtaining an audio quality almost on par to single-precision DNNs and a slightly larger Word Error Rate (WER) for single speaker scenarios using the WSJ0 speech corpus.

Radar sensors are crucial for environment perception of driver assistance systems as well as autonomous vehicles. Key performance factors are weather resistance and the possibility to directly measure velocity. With a rising number of radar sensors and the so far unregulated automotive radar frequency band, mutual interference is inevitable and must be dealt with. Algorithms and models operating on radar data in early processing stages are required to run directly on specialized hardware, i.e. the radar sensor. This specialized hardware typically has strict resource-constraints, i.e. a low memory capacity and low computational power. Convolutional Neural Network (CNN)-based approaches for denoising and interference mitigation yield promising results for radar processing in terms of performance. However, these models typically contain millions of parameters, stored in hundreds of megabytes of memory, and require additional memory during execution. In this paper we investigate quantization techniques for CNN-based denoising and interference mitigation of radar signals. We analyze the quantization potential of different CNN-based model architectures and sizes by considering (i) quantized weights and (ii) piecewise constant activation functions, which results in reduced memory requirements for model storage and during the inference step respectively.

Since resource-constrained devices hardly benefit from the trend towards ever-increasing neural network (NN) structures, there is growing interest in designing more hardware-friendly NNs. In this paper, we consider the training of NNs with discrete-valued weights and sign activation functions that can be implemented more efficiently in terms of inference speed, memory requirements, and power consumption. We build on the framework of probabilistic forward propagations using the local reparameterization trick, where instead of training a single set of NN weights we rather train a distribution over these weights. Using this approach, we can perform gradient-based learning by optimizing the continuous distribution parameters over discrete weights while at the same time perform backpropagation through the sign activation. In our experiments, we investigate the influence of the number of weights on the classification performance on several benchmark datasets, and we show that our method achieves state-of-the-art performance.

We extend feed-forward neural networks with a Dirichlet process prior over the weight distribution. This enforces a sharing on the network weights, which can reduce the overall number of parameters drastically. We alternately sample from the posterior of the weights and the posterior of assignments of network connections to the weights. This results in a weight sharing that is adopted to the given data. In order to make the procedure feasible, we present several techniques to reduce the computational burden. Experiments show that our approach mostly outperforms models with random weight sharing. Our model is capable of reducing the memory footprint substantially while maintaining a good performance compared to neural networks without weight sharing.

Learning the structure of Bayesian networks is a difficult combinatorial optimization problem. In this paper, we consider learning of tree-augmented naive Bayes (TAN) structures for Bayesian network classifiers with discrete input features. Instead of performing a combinatorial optimization over the space of possible graph structures, the proposed method learns a distribution over graph structures. After training, we select the most probable structure of this distribution. This allows for a joint training of the Bayesian network parameters along with its TAN structure using gradient-based optimization. The proposed method is agnostic to the specific loss and only requires that it is differentiable. We perform extensive experiments using a hybrid generative-discriminative loss based on the discriminative probabilistic margin. Our method consistently outperforms random TAN structures and Chow-Liu TAN structures.

Modern musical source separation systems based on deep neural networks reach unprecedented levels of separation quality. However, harnessing the power of these large-scale models in typical audio production environments, which frequently offer only limited computing resources while demanding real-time processing, remains challenging. We extend the multi-scaled DenseNet in several aspects to facilitate real-time source separation scenarios. Specifically, we reduce the computational requirements by inferring Melscaled masks and decrease the model size via effective use of bottleneck layers, while improving performance using a deep clustering objective. In addition, we are able to further increase the model efficiency by applying parameterized structured pruning of convolutional weights without any significant impact on the separation performance. We significantly reduce the model size and increase the computational efficiency by a factor of 1.6 and 4.3, respectively, while maintaining the separation performance.

Models play an essential role in the design process of cyber-physical systems. They form the basis for simulation and analysis and help in identifying design problems as early as possible. However, the construction of models that comprise physical and digital behavior is challenging. Therefore, there is considerable interest in learning such hybrid behavior by means of machine learning which requires sufficient and representative training data covering the behavior of the physical system adequately. In this work, we exploit a combination of automata learning and model-based testing to generate sufficient training data fully automatically. Experimental results on a platooning scenario show that recurrent neural networks learned with this data achieved significantly better results compared to models learned from randomly generated data. In particular, the classification error for crash detection is reduced by a factor of five and a similar F1-score is obtained with up to three orders of magnitude fewer training samples.

The tremendous memory and computational complexity of Convolutional Neural Networks (CNNs) prevents the inference deployment on resource-constrained systems. As a result, recent research focused on CNN optimization techniques, in particular quantization, which allows weights and activations of layers to be represented with just a few bits while achieving impressive prediction performance. However, aggressive quantization techniques still fail to achieve full-precision prediction performance on state-of-the-art CNN architectures on large-scale classification tasks. In this work we propose a method for weight and activation quantization that is scalable in terms of quantization levels (n-ary representations) and easy to compute while maintaining the performance close to full-precision CNNs. Our weight quantization scheme is based on trainable scaling factors and a nested-means clustering strategy which is robust to weight updates and therefore exhibits good convergence properties. The flexibility of nested-means clustering enables exploration of various n-ary weight representations with the potential of high parameter compression. For activations, we propose a linear quantization strategy that takes the statistical properties of batch normalization into account. We demonstrate the effectiveness of our approach using state-of-the-art models on ImageNet.

Recent work has shown substantial performance improvements of discriminative probabilistic models over their generative counterparts. However, since discriminative models do not capture the input distribution of the data, their use in missing data scenarios is limited. To utilize the advantages of both paradigms, we present an approach to train Gaussian mixture models (GMMs) in a hybrid generative-discriminative way. This is accomplished by optimizing an objective that trades off between a generative likelihood term and either a discriminative conditional likelihood term or a large margin term using stochastic optimization. Our model substantially improves the performance of classical maximum likelihood optimized GMMs while at the same time allowing for both a consistent treatment of missing features by marginalization, and the use of additional unlabeled data in a semi-supervised setting. For the covariance matrices, we employ a diagonal plus low-rank matrix structure to model important correlations while keeping the number of parameters small. We show that a non-diagonal matrix structure is crucial to achieve good performance and that the proposed structure can be utilized to considerably reduce classification time in case of missing features. The capabilities of our model are demonstrated in extensive experiments on real-world data. (c) 2018 Published by Elsevier B V.

The increasing demand for neural networks (NNs) being employed on embedded devices has led to plenty of research investigating methods for training low precision NNs. While most methods involve a quantization step, we propose a principled Bayesian approach where we first infer a distribution over a discrete weight space from which we subsequently derive hardware-friendly low precision NNs. To this end, we introduce a probabilistic forward pass to approximate the intractable variational objective that allows us to optimize over discrete-valued weight distributions for NNs with sign activation functions. In our experiments, we show that our model achieves state of the art performance on several real world data sets. In addition, the resulting models exhibit a substantial amount of sparsity that can be utilized to further reduce the computational costs for inference.

While machine learning is traditionally a resource intensive task, embedded systems, autonomous navigation and the vision of the Internet-of-Things fuel the interest in resource efficient approaches. These approaches require a carefully chosen trade-off between performance and resource consumption in terms of computation and energy. On top of this, it is crucial to treat uncertainty in a consistent manner in all but the simplest applications of machine learning systems. In particular, a desideratum for any real-world system is to be robust in the presence of outliers and corrupted data, as well as being `aware' of its limits, i.e. the system should maintain and provide an uncertainty estimate over its own predictions. These complex demands are among the major challenges in current machine learning research and key to ensure a smooth transition of machine learning technology into every day's applications. In this article, we provide an overview of the current state of the art of machine learning techniques facilitating these real-world requirements. First we provide a comprehensive review of resource-efficiency in deep neural networks with focus on techniques for model size reduction, compression and reduced precision. These techniques can be applied during training or as post-processing and are widely used to reduce both computational complexity and memory footprint. As most (practical) neural networks are limited in their ways to treat uncertainty, we contrast them with probabilistic graphical models, which readily serve these desiderata by means of probabilistic inference. In that way, we provide an extensive overview of the current state-of-the-art of robust and efficient machine learning for real-world systems.

In recent years, voice activity detection has been a highly researched field, due to its importance as input stage in many real-world applications. Automated detection of vocalisations in the very first year of life is still a stepchild of this field. On our quest defining acoustic parameters in pre-linguistic vocalisations as markers for neuro(mal)development, we are confronted with the challenge of manually segmenting and annotating hours of variable quality home video material for sequences of infant voice/vocalisations. While in total our corpus comprises video footage of typically developing infants and infants with various neurodevelopmental disorders of more than a year running time, only a small proportion has been processed so far. This calls for automated assistance tools for detecting and/or segmenting infant utterances from real-live video recordings. In this paper, we investigated several approaches of infant voice detection and segmentation, including a rule-based voice activity detector, hidden Markov models with Gaussian mixture observation models, support vector machines, and random forests. Results indicate that the applied methods could be well applied in a semi-automated retrieval of infant utterances from highly non-standardised footage. At the same time, our results show that, a fully automated approach for this problem is yet to come.