Search

Matrix operations are nowadays central in many Machine Learning techniques, including in particular Deep Neural Networks (DNNs), whose core of any inference is represented by a sequence of dot product operations. An increasingly emerging problem is how to efficiently engineer their storage and operations. In this article we propose two new lossless compression schemes for real-valued matrices, supporting efficient vector-matrix multiplications in the compressed format, and specifically suitable for DNNs compression. Exploiting several recent studies that use weight pruning and quantization techniques to reduce the complexity of DNN inference, our schemes are expressly designed to benefit from both, that is from input matrices characterized by low entropy. In particular, our solutions are able to take advantage from the depth of the model, and the deeper the model, the higher the efficiency. Moreover, we derived space upper bounds for both variants in terms of the source entropy. Experiments show that our tools favourably compare in terms of energy and space efficiency against state-of-the-art matrix compression approaches, including Compressed Linear Algebra (CLA) and Compressed Shared Elements Row (CSER), the latter explicitly proposed in the context of DNN compression.

Multiobject detection has become an integral component in various neural applications, such as autonomous driving and augmented reality. The system should be able to recognize and process multiple objects simultaneously. Moreover, the performance requirements for this system can be dynamically changed depending on the number of regions of interest (ROIs) in each frame. Consequently, the processing unit (PU) of the neural acceleration system should provide various inference rates. Therefore, we present a field-programmable gate array (FPGA)-based dynamic rate neural acceleration system called MultiLockOn to dynamically change the inference performance according to the number of ROIs per frame. It supports multiprocessing modes with different speeds through the introduction of novel multi-mode processing engines (PEs) comprising minimum reconfigurable interconnections across inference modes to minimize hardware overhead. The MultiLockOn system can provide an improvement of up to $4\times $ in the inference performance compared to that of DNNWeaver and $5.7\times $ compared to that of the ARM Cortex-A53 with minimum accuracy loss by supporting the multiprocessing modes. [ABSTRACT FROM AUTHOR]

Recently, various memory devices have been actively studied as suitable candidates for synaptic devices, which are important memory and computing units in neuromorphic systems. One of the ways to manage these devices is off-chip training, where it is essential to transfer the pretrained weights accurately. Previous studies, however, have a few limitations, such as a lack of consideration of program errors that occur during the transfer process. Although the smaller the program error, the higher the accuracy, the corresponding increase in the program time must be considered. To evaluate the practical applicability, we fabricated Al2O3/TiOx-based resistive random access memory (RRAM) and investigated the effect of program errors on program time and system degradation. It was confirmed that for smaller program errors, the program time was exponentially longer. Furthermore, we examined the effect of variation with respect to the number of quantized weight states (${N}_{state}$) through system-level simulation. We observed that the optimized ${N}_{state}$ varies depending on whether the program error is small or large. This result is meaningful as it experimentally shows the tradeoff between the program error, program time, and system performance. We expect it to be useful in the development of neuromorphic systems. [ABSTRACT FROM AUTHOR]

With the refinement of tasks in artificial intelligence, bringing in exponential level increments in computation cost and storage. Therefore, the augment of computation resource for complicated neural networks severely hinders their applications on limited-power devices in recent years. As a result, there is an impending necessity to compress and accelerate the deep networks by special ways. Considering the different peculiarities of weight quantization and sparse regularization, in this paper, we propose a low rank sparse quantization (LRSQ) method to quantize network weights and regularize the corresponding structures at the same time. Our LRSQ can: 1) obtain low-bit quantized networks to reduce memory and computation cost and 2) learn a compact structure from complex convolutional networks for subsequent channel pruning which has significant reduction on FLOPs. In experimental sections, we evaluate the proposed method on several popular models such as VGG-7/16/19 and ResNet-18/34/50, and results show that this method can dramatically reduce parameters and channels of the network with slight inference accuracy loss. Furthermore, we also visualize and analyze the four-dimensional weight tensors, which shows the low rank and group-sparsity structure of it. Finally, we try pruning unimportant channels which are zero-channels in our quantized model, and finding even a little better precision than the standard full-precision network.

The primary objective of this paper is to build classification models and strategies to identify breathing sound anomalies (wheeze, crackle) for automated diagnosis of respiratory and pulmonary diseases. In this work we propose a deep CNN-RNN model that classifies respiratory sounds based on Mel-spectrograms. We also implement a patient specific model tuning strategy that first screens respiratory patients and then builds patient specific classification models using limited patient data for reliable anomaly detection. Moreover, we devise a local log quantization strategy for model weights to reduce the memory footprint for deployment in memory constrained systems such as wearable devices. The proposed hybrid CNN-RNN model achieves a score of $\text{66.31}\%$ on four-class classification of breathing cycles for ICBHI’17 scientific challenge respiratory sound database. When the model is re-trained with patient specific data, it produces a score of $\text{71.81}\%$ for leave-one-out validation. The proposed weight quantization technique achieves ${\approx}4 \times$ reduction in total memory cost without loss of performance. The main contribution of the paper is as follows: Firstly, the proposed model is able to achieve state of the art score on the ICBHI’17 dataset. Secondly, deep learning models are shown to successfully learn domain specific knowledge when pre-trained with breathing data and produce significantly superior performance compared to generalized models. Finally, local log quantization of trained weights is shown to be able to reduce the memory requirement significantly. This type of patient-specific re-training strategy can be very useful in developing reliable long-term automated patient monitoring systems particularly in wearable healthcare solutions. [ABSTRACT FROM AUTHOR]

Deep neural networks enable state-of-the-art accuracy on visual recognition tasks such as image classification and object detection. However, modern networks contain millions of learned connections, and the current trend is towards deeper and more densely connected architectures. This poses a challenge to the deployment of state-of-the-art networks on resource-constrained systems, such as smartphones or mobile robots. In general, a more efficient utilization of computation resources would assist in deployment scenarios from embedded platforms to computing clusters running ensembles of networks. In this paper, we propose a deep network compression algorithm that performs weight pruning and quantization jointly, and in parallel with fine-tuning. Our approach takes advantage of the complementary nature of pruning and quantization and recovers from premature pruning errors, which is not possible with two-stage approaches. In experiments on ImageNet, CLIP-Q (Compression Learning by In-Parallel Pruning-Quantization) improves the state-of-the-art in network compression on AlexNet, VGGNet, GoogLeNet, and ResNet. We additionally demonstrate that CLIP-Q is complementary to efficient network architecture design by compressing MobileNet and ShuffleNet, and that CLIP-Q generalizes beyond convolutional networks by compressing a memory network for visual question answering. [ABSTRACT FROM AUTHOR]

Due to controllable conductance and non-volatility, flexible memristors are regarded as a key enabler for building artificial neural network (ANN)-based learning algorithms in flexible and wearable systems. However, the existing flexible memristors are suffering from limited number of conductance values, issues limiting large-scale integration, and insufficient accuracy that cannot support accurate computation of ANN. In this paper, solutions are proposed for the three major challenges of the flexible memristor; the feasibility of a three-layer fully connected neural network on MNIST and a 13-layer convolutional neural network (CNN) on CIFAR-10 using the flexible memristor based on single-walled carbon nanotubes network/polymer composite and hydrophilic Al2O3 dielectric are studied. The evaluation result shows that in the fully connected neural network system, it is able to recognize MNIST with an accuracy above 90% after 4-bit quantization, 52.05% decrease in interconnection numbers in the circuit and up to 40% random error introduced, and in the CNN on CIFAR-10, the system can retain an accuracy above 86% with less than 4% accuracy loss after 5-bit quantization, 59.34% decrease in interconnection numbers in the circuit and up to 40% random error injected.

Spiking neural networks (SNNs) with a large number of weights and varied weight distribution can be difficult to implement in emerging in-memory computing hardware due to the limitations on crossbar size (implementing dot product), the constrained number of conductance states in non-CMOS devices and the power budget. We present a sparse SNN topology where noncritical connections are pruned to reduce the network size, and the remaining critical synapses are weight quantized to accommodate for limited conductance states. Pruning is based on the power law weight-dependent spike timing dependent plasticity model; synapses between pre- and post-neuron with high spike correlation are retained, whereas synapses with low correlation or uncorrelated spiking activity are pruned. The weights of the retained connections are quantized to the available number of conductance states. The process of pruning noncritical connections and quantizing the weights of critical synapses is performed at regular intervals during training. We evaluated our sparse and quantized network on MNIST dataset and on a subset of images from Caltech-101 dataset. The compressed topology achieved a classification accuracy of 90.1% (91.6%) on the MNIST (Caltech-101) dataset with 3.1X (2.2X) and 4X (2.6X) improvement in energy and area, respectively. The compressed topology is energy and area efficient while maintaining the same classification accuracy of a 2-layer fully connected SNN topology. [ABSTRACT FROM AUTHOR]