We present a computationally efficient method for compressing a trained neural network without using real data. We break the problem of data-free network compression into independent layer-wise compressions. We show how to efficiently generate layer-wise training data using only a pretrained network. We use this data to perform independent layer-wise compressions on the pretrained network. We also show how to precondition the network to improve the accuracy of our layer-wise compression method. We present results for layer-wise compression using quantization and pruning. When quantizing, we compress with higher accuracy than related works while using orders of magnitude less compute. When compressing MobileNetV2 and evaluating on ImageNet, our method outperforms existing methods for quantization at all bit-widths, achieving a $+0.34\%$ improvement in $8$-bit quantization, and a stronger improvement at lower bit-widths (up to a $+28.50\%$ improvement at $5$ bits). When pruning, we outperform baselines of a similar compute envelope, achieving $1.5$ times the sparsity rate at the same accuracy. We also show how to combine our efficient method with high-compute generative methods to improve upon their results.
Time series forecasting is essential for agents to make decisions. Traditional approaches rely on statistical methods to forecast given past numeric values. In practice, end-users often rely on visualizations such as charts and plots to reason about their forecasts. Inspired by practitioners, we re-imagine the topic by creating a novel framework to produce visual forecasts, similar to the way humans intuitively do. In this work, we leverage advances in deep learning to extend the field of time series forecasting to a visual setting. We capture input data as an image and train a model to produce the subsequent image. This approach results in predicting distributions as opposed to pointwise values. We examine various synthetic and real datasets with diverse degrees of complexity. Our experiments show that visual forecasting is effective for cyclic data but somewhat less for irregular data such as stock price. Importantly, when using image-based evaluation metrics, we find the proposed visual forecasting method to outperform various numerical baselines, including ARIMA and a numerical variation of our method. We demonstrate the benefits of incorporating vision-based approaches in forecasting tasks -- both for the quality of the forecasts produced, as well as the metrics that can be used to evaluate them.
Quantization for deep neural networks (DNN) have enabled developers to deploy models with less memory and more efficient low-power inference. However, not all DNN designs are friendly to quantization. For example, the popular Mobilenet architecture has been tuned to reduce parameter size and computational latency with separable depth-wise convolutions, but not all quantization algorithms work well and the accuracy can suffer against its float point versions. In this paper, we analyzed several root causes of quantization loss and proposed alternatives that do not rely on per-channel or training-aware approaches. We evaluate the image classification task on ImageNet dataset, and our post-training quantized 8-bit inference top-1 accuracy in within 0.7% of the floating point version.
Deep Neural Networks(DNNs) have many parameters and activation data, and these both are expensive to implement. One method to reduce the size of the DNN is to quantize the pre-trained model by using a low-bit expression for weights and activations, using fine-tuning to recover the drop in accuracy. However, it is generally difficult to train neural networks which use low-bit expressions. One reason is that the weights in the middle layer of the DNN have a wide dynamic range and so when quantizing the wide dynamic range into a few bits, the step size becomes large, which leads to a large quantization error and finally a large degradation in accuracy. To solve this problem, this paper makes the following three contributions without using any additional learning parameters and hyper-parameters. First, we analyze how batch normalization, which causes the aforementioned problem, disturbs the fine-tuning of the quantized DNN. Second, based on these results, we propose a new pruning method called Pruning for Quantization (PfQ) which removes the filters that disturb the fine-tuning of the DNN while not affecting the inferred result as far as possible. Third, we propose a workflow of fine-tuning for quantized DNNs using the proposed pruning method(PfQ). Experiments using well-known models and datasets confirmed that the proposed method achieves higher performance with a similar model size than conventional quantization methods including fine-tuning.
In the traditional deep compression framework, iteratively performing network pruning and quantization can reduce the model size and computation cost to meet the deployment requirements. However, such a step-wise application of pruning and quantization may lead to suboptimal solutions and unnecessary time consumption. In this paper, we tackle this issue by integrating network pruning and quantization as a unified joint compression problem and then use AutoML to automatically solve it. We find the pruning process can be regarded as the channel-wise quantization with 0 bit. Thus, the separate two-step pruning and quantization can be simplified as the one-step quantization with mixed precision. This unification not only simplifies the compression pipeline but also avoids the compression divergence. To implement this idea, we propose the automated model compression by jointly applied pruning and quantization (AJPQ). AJPQ is designed with a hierarchical architecture: the layer controller controls the layer sparsity, and the channel controller decides the bit-width for each kernel. Following the same importance criterion, the layer controller and the channel controller collaboratively decide the compression strategy. With the help of reinforcement learning, our one-step compression is automatically achieved. Compared with the state-of-the-art automated compression methods, our method obtains a better accuracy while reducing the storage considerably. For fixed precision quantization, AJPQ can reduce more than five times model size and two times computation with a slight performance increase for Skynet in remote sensing object detection. When mixed-precision is allowed, AJPQ can reduce five times model size with only 1.06% top-5 accuracy decline for MobileNet in the classification task.
Cellular and molecular imaging techniques and models have been developed to characterize single stages of viral proliferation after focal infection of cells in vitro. The fast and automatic classification of cell imaging data may prove helpful prior to any further comparison of representative experimental data to mathematical models of viral propagation in host cells. Here, we use computer generated images drawn from a reproduction of an imaging model from a previously published study of experimentally obtained cell imaging data representing progressive viral particle proliferation in host cell monolayers. Inspired by experimental time-based imaging data, here in this study viral particle increase in time is simulated by a one-by-one increase, across images, in black or gray single pixels representing dead or partially infected cells, and hypothetical remission by a one-by-one increase in white pixels coding for living cells in the original image model. The image simulations are submitted to unsupervised learning by a Self-Organizing Map (SOM) and the Quantization Error in the SOM output (SOM-QE) is used for automatic classification of the image simulations as a function of the represented extent of viral particle proliferation or cell recovery. Unsupervised classification by SOM-QE of 160 model images, each with more than three million pixels, is shown to provide a statistically reliable, pixel precise, and fast classification model that outperforms human computer-assisted image classification by RGB image mean computation. The automatic classification procedure proposed here provides a powerful approach to understand finely tuned mechanisms in the infection and proliferation of virus in cell lines in vitro or other cells.
The quantization error in a fixed-size Self-Organizing Map (SOM) with unsupervised winner-take-all learning has previously been used successfully to detect, in minimal computation time, highly meaningful changes across images in medical time series and in time series of satellite images. Here, the functional properties of the quantization error in SOM are explored further to show that the metric is capable of reliably discriminating between the finest differences in local contrast intensities and contrast signs. While this capability of the QE is akin to functional characteristics of a specific class of retinal ganglion cells (the so-called Y-cells) in the visual systems of the primate and the cat, the sensitivity of the QE surpasses the capacity limits of human visual detection. Here, the quantization error in the SOM is found to reliably signal changes in contrast or colour when contrast information is removed from or added to the image, but not when the amount and relative weight of contrast information is constant and only the local spatial position of contrast elements in the pattern changes. While the RGB Mean reflects coarser changes in colour or contrast well enough, the SOM-QE is shown to outperform the RGB Mean in the detection of single-pixel changes in images with up to five million pixels. This could have important implications in the context of unsupervised image learning and computational building block approaches to large sets of image data (big data), including deep learning blocks, and automatic detection of contrast change at the nanoscale in Transmission or Scanning Electron Micrographs (TEM, SEM), or at the subpixel level in multispectral and hyper-spectral imaging data.
Deep convolutional neural networks (DCNNs) have shown dominant performance in the task of super-resolution (SR). However, their heavy memory cost and computation overhead significantly restrict their practical deployments on resource-limited devices, which mainly arise from the floating-point storage and operations between weights and activations. Although previous endeavors mainly resort to fixed-point operations, quantizing both weights and activations with fixed coding lengths may cause significant performance drop, especially on low bits. Specifically, most state-of-the-art SR models without batch normalization have a large dynamic quantization range, which also serves as another cause of performance drop. To address these two issues, we propose a new quantization scheme termed PArameterized Max Scale (PAMS), which applies the trainable truncated parameter to explore the upper bound of the quantization range adaptively. Finally, a structured knowledge transfer (SKT) loss is introduced to fine-tune the quantized network. Extensive experiments demonstrate that the proposed PAMS scheme can well compress and accelerate the existing SR models such as EDSR and RDN. Notably, 8-bit PAMS-EDSR improves PSNR on Set5 benchmark from 32.095dB to 32.124dB with 2.42$\times$ compression ratio, which achieves a new state-of-the-art.
Voice user interfaces and digital assistants are rapidly entering our lives and becoming singular touch points spanning our devices. These always-on services capture and transmit our audio data to powerful cloud services for further processing and subsequent actions. Our voices and raw audio signals collected through these devices contain a host of sensitive paralinguistic information that is transmitted to service providers regardless of deliberate or false triggers. As our emotional patterns and sensitive attributes like our identity, gender, well-being, are easily inferred using deep acoustic models, we encounter a new generation of privacy risks by using these services. One approach to mitigate the risk of paralinguistic-based privacy breaches is to exploit a combination of cloud-based processing with privacy-preserving, on-device paralinguistic information learning and filtering before transmitting voice data. In this paper we introduce EDGY, a configurable, lightweight, disentangled representation learning framework that transforms and filters high-dimensional voice data to identify and contain sensitive attributes at the edge prior to offloading to the cloud. We evaluate EDGY's on-device performance and explore optimization techniques, including model quantization and knowledge distillation, to enable private, accurate and efficient representation learning on resource-constrained devices. Our results show that EDGY runs in tens of milliseconds with 0.2% relative improvement in "zero-shot" ABX score or minimal performance penalties of approximately 5.95% word error rate (WER) in learning linguistic representations from raw voice signals, using a CPU and a single-core ARM processor without specialized hardware.
Building footprint extraction (BFE) from multi-sensor data such as optical images and light detection and ranging (LiDAR) point clouds is widely used in various fields of remote sensing applications. However, it is still challenging research topic due to relatively inefficient building extraction techniques from variety of complex scenes in multi-sensor data. In this study, we develop and evaluate a deep competition network (DCN) that fuses very high spatial resolution optical remote sensing images with LiDAR data for robust BFE. DCN is a deep superpixelwise convolutional encoder-decoder architecture using the encoder vector quantization with classified structure. DCN consists of five encoding-decoding blocks with convolutional weights for robust binary representation (superpixel) learning. DCN is trained and tested in a big multi-sensor dataset obtained from the state of Indiana in the United States with multiple building scenes. Comparison results of the accuracy assessment showed that DCN has competitive BFE performance in comparison with other deep semantic binary segmentation architectures. Therefore, we conclude that the proposed model is a suitable solution to the robust BFE from big multi-sensor data.
Real-time understanding in video is crucial in various AI applications such as autonomous driving. This work presents a fast single-shot segmentation strategy for video scene understanding. The proposed net, called S3-Net, quickly locates and segments target sub-scenes, meanwhile extracts structured time-series semantic features as inputs to an LSTM-based spatio-temporal model. Utilizing tensorization and quantization techniques, S3-Net is intended to be lightweight for edge computing. Experiments using CityScapes, UCF11, HMDB51 and MOMENTS datasets demonstrate that the proposed S3-Net achieves an accuracy improvement of 8.1% versus the 3D-CNN based approach on UCF11, a storage reduction of 6.9x and an inference speed of 22.8 FPS on CityScapes with a GTX1080Ti GPU.
Research in computational deep learning has directed considerable efforts towards hardware-oriented optimisations for deep neural networks, via the simplification of the activation functions, or the quantization of both activations and weights. The resulting non-differentiability (or even discontinuity) of the networks poses some challenging problems, especially in connection with the learning process. In this paper, we address several questions regarding both the expressivity of quantized neural networks and approximation techniques for non-differentiable networks. First, we answer in the affirmative the question of whether QNNs have the same expressivity as DNNs in terms of approximation of Lipschitz functions in the $L^{\infty}$ norm. Then, considering a continuous but not necessarily differentiable network, we describe a layer-wise stochastic regularisation technique to produce differentiable approximations, and we show how this approach to regularisation provides elegant quantitative estimates. Finally, we consider networks defined by means of Heaviside-type activation functions, and prove for them a pointwise approximation result by means of smooth networks under suitable assumptions on the regularised activations.
Decentralized optimization methods enable on-device training of machine learning models without a central coordinator. In many scenarios communication between devices is energy demanding and time consuming and forms the bottleneck of the entire system. We propose a new randomized first-order method which tackles the communication bottleneck by applying randomized compression operators to the communicated messages. By combining our scheme with a new variance reduction technique that progressively throughout the iterations reduces the adverse effect of the injected quantization noise, we obtain the first scheme that converges linearly on strongly convex decentralized problems while using compressed communication only. We prove that our method can solve the problems without any increase in the number of communications compared to the baseline which does not perform any communication compression while still allowing for a significant compression factor which depends on the conditioning of the problem and the topology of the network. Our key theoretical findings are supported by numerical experiments.
Distributed model training suffers from communication bottlenecks due to frequent model updates transmitted across compute nodes. To alleviate these bottlenecks, practitioners use gradient compression techniques like sparsification, quantization, or low-rank updates. The techniques usually require choosing a static compression ratio, often requiring users to balance the trade-off between model accuracy and per-iteration speedup. In this work, we show that such performance degradation due to choosing a high compression ratio is not fundamental. An adaptive compression strategy can reduce communication while maintaining final test accuracy. Inspired by recent findings on critical learning regimes, in which small gradient errors can have irrecoverable impact on model performance, we propose Accordion a simple yet effective adaptive compression algorithm. While Accordion maintains a high enough compression rate on average, it avoids over-compressing gradients whenever in critical learning regimes, detected by a simple gradient-norm based criterion. Our extensive experimental study over a number of machine learning tasks in distributed environments indicates that Accordion, maintains similar model accuracy to uncompressed training, yet achieves up to 5.5x better compression and up to 4.1x end-to-end speedup over static approaches. We show that Accordion also works for adjusting the batch size, another popular strategy for alleviating communication bottlenecks.
We propose a new computationally efficient method for quantizing the weights of pre- trained neural networks that is general enough to handle both multi-layer perceptrons and convolutional neural networks. Our method deterministically quantizes layers in an iterative fashion with no complicated re-training required. Specifically, we quantize each neuron, or hidden unit, using a greedy path-following algorithm. This simple algorithm is equivalent to running a dynamical system, which we prove is stable for quantizing a single-layer neural network (or, alternatively, for quantizing the first layer of a multi-layer network) when the training data are Gaussian. We show that under these assumptions, the quantization error decays with the width of the layer, i.e., its level of over-parametrization. We provide numerical experiments, on multi-layer networks, to illustrate the performance of our methods on MNIST and CIFAR10 data, as well as for quantizing the VGG16 network using ImageNet data.
Recent advances in acquisition equipment is providing experiments with growing amounts of precise yet affordable sensors. At the same time an improved computational power, coming from new hardware resources (GPU, FPGA, ACAP), has been made available at relatively low costs. This led us to explore the possibility of completely renewing the chain of acquisition for a fusion experiment, where many high-rate sources of data, coming from different diagnostics, can be combined in a wide framework of algorithms. If on one hand adding new data sources with different diagnostics enriches our knowledge about physical aspects, on the other hand the dimensions of the overall model grow, making relations among variables more and more opaque. A new approach for the integration of such heterogeneous diagnostics, based on composition of deep variational autoencoders, could ease this problem, acting as a structural sparse regularizer. This has been applied to RFX-mod experiment data, integrating the soft X-ray linear images of plasma temperature with the magnetic state. However to ensure a real-time signal analysis, those algorithmic techniques must be adapted to run in well suited hardware. In particular it is shown that, attempting a quantization of neurons transfer functions, such models can be modified to create an embedded firmware. This firmware, approximating the deep inference model to a set of simple operations, fits well with the simple logic units that are largely abundant in FPGAs. This is the key factor that permits the use of affordable hardware with complex deep neural topology and operates them in real-time.
Compressing large neural networks is an important step for their deployment in resource-constrained computational platforms. In this context, vector quantization is an appealing framework that expresses multiple parameters using a single code, and has recently achieved state-of-the-art network compression on a range of core vision and natural language processing tasks. Key to the success of vector quantization is deciding which parameter groups should be compressed together. Previous work has relied on heuristics that group the spatial dimension of individual convolutional filters, but a general solution remains unaddressed. This is desirable for pointwise convolutions (which dominate modern architectures), linear layers (which have no notion of spatial dimension), and convolutions (when more than one filter is compressed to the same codeword). In this paper we make the observation that the weights of two adjacent layers can be permuted while expressing the same function. We then establish a connection to rate-distortion theory and search for permutations that result in networks that are easier to compress. Finally, we rely on an annealed quantization algorithm to better compress the network and achieve higher final accuracy. We show results on image classification, object detection, and segmentation, reducing the gap with the uncompressed model by 40 to 70% with respect to the current state of the art.
The intensive computation of Automatic Speech Recognition (ASR) models obstructs them from being deployed on mobile devices. In this paper, we present a novel quantized Winograd optimization pipeline, which combines the quantization and fast convolution to achieve efficient inference acceleration on mobile devices for ASR models. To avoid the information loss due to the combination of quantization and Winograd convolution, a Range-Scaled Quantization (RSQ) training method is proposed to expand the quantized numerical range and to distill knowledge from high-precision values. Moreover, an improved Conv1D equipped DFSMN (ConvDFSMN) model is designed for mobile deployment. We conduct extensive experiments on both ConvDFSMN and Wav2letter models. Results demonstrate the models can be effectively optimized with the proposed pipeline. Especially, Wav2letter achieves 1.48* speedup with an approximate 0.07% WER decrease on ARMv7-based mobile devices.
Convolutional neural networks (CNNs) play a key role in deep learning applications. However, the large storage overheads and the substantial computation cost of CNNs are problematic in hardware accelerators. Computing-in-memory (CIM) architecture has demonstrated great potential to effectively compute large-scale matrix-vector multiplication. However, the intensive multiply and accumulation (MAC) operations executed at the crossbar array and the limited capacity of CIM macros remain bottlenecks for further improvement of energy efficiency and throughput. To reduce computation costs, network pruning and quantization are two widely studied compression methods to shrink the model size. However, most of the model compression algorithms can only be implemented in digital-based CNN accelerators. For implementation in a static random access memory (SRAM) CIM-based accelerator, the model compression algorithm must consider the hardware limitations of CIM macros, such as the number of word lines and bit lines that can be turned on at the same time, as well as how to map the weight to the SRAM CIM macro. In this study, a software and hardware co-design approach is proposed to design an SRAM CIM-based CNN accelerator and an SRAM CIM-aware model compression algorithm. To lessen the high-precision MAC required by batch normalization (BN), a quantization algorithm that can fuse BN into the weights is proposed. Furthermore, to reduce the number of network parameters, a sparsity algorithm that considers a CIM architecture is proposed. Last, MARS, a CIM-based CNN accelerator that can utilize multiple SRAM CIM macros as processing units and support a sparsity neural network, is proposed.
Multiplication (e.g., convolution) is arguably a cornerstone of modern deep neural networks (DNNs). However, intensive multiplications cause expensive resource costs that challenge DNNs' deployment on resource-constrained edge devices, driving several attempts for multiplication-less deep networks. This paper presented ShiftAddNet, whose main inspiration is drawn from a common practice in energy-efficient hardware implementation, that is, multiplication can be instead performed with additions and logical bit-shifts. We leverage this idea to explicitly parameterize deep networks in this way, yielding a new type of deep network that involves only bit-shift and additive weight layers. This hardware-inspired ShiftAddNet immediately leads to both energy-efficient inference and training, without compromising the expressive capacity compared to standard DNNs. The two complementary operation types (bit-shift and add) additionally enable finer-grained control of the model's learning capacity, leading to more flexible trade-off between accuracy and (training) efficiency, as well as improved robustness to quantization and pruning. We conduct extensive experiments and ablation studies, all backed up by our FPGA-based ShiftAddNet implementation and energy measurements. Compared to existing DNNs or other multiplication-less models, ShiftAddNet aggressively reduces over 80% hardware-quantified energy cost of DNNs training and inference, while offering comparable or better accuracies. Codes and pre-trained models are available at https://github.com/RICE-EIC/ShiftAddNet.