We present Random Partition Relaxation (RPR), a method for strong quantization of neural networks weight to binary (+1/-1) and ternary (+1/0/-1) values. Starting from a pre-trained model, we quantize the weights and then relax random partitions of them to their continuous values for retraining before re-quantizing them and switching to another weight partition for further adaptation. We demonstrate binary and ternary-weight networks with accuracies beyond the state-of-the-art for GoogLeNet and competitive performance for ResNet-18 and ResNet-50 using an SGD-based training method that can easily be integrated into existing frameworks.
In this paper, we present a new on-device automatic speech recognition (ASR) system based on monotonic chunk-wise attention (MoChA) models trained with large (> 10K hours) corpus. We attained around 90% of a word recognition rate for general domain mainly by using joint training of connectionist temporal classifier (CTC) and cross entropy (CE) losses, minimum word error rate (MWER) training, layer-wise pre-training and data augmentation methods. In addition, we compressed our models by more than 3.4 times smaller using an iterative hyper low-rank approximation (LRA) method while minimizing the degradation in recognition accuracy. The memory footprint was further reduced with 8-bit quantization to bring down the final model size to lower than 39 MB. For on-demand adaptation, we fused the MoChA models with statistical n-gram models, and we could achieve a relatively 36% improvement on average in word error rate (WER) for target domains including the general domain.
While increasingly deep networks are still in general desired for achieving state-of-the-art performance, for many specific inputs a simpler network might already suffice. Existing works exploited this observation by learning to skip convolutional layers in an input-dependent manner. However, we argue their binary decision scheme, i.e., either fully executing or completely bypassing one layer for a specific input, can be enhanced by introducing finer-grained, "softer" decisions. We therefore propose a Dynamic Fractional Skipping (DFS) framework. The core idea of DFS is to hypothesize layer-wise quantization (to different bitwidths) as intermediate "soft" choices to be made between fully utilizing and skipping a layer. For each input, DFS dynamically assigns a bitwidth to both weights and activations of each layer, where fully executing and skipping could be viewed as two "extremes" (i.e., full bitwidth and zero bitwidth). In this way, DFS can "fractionally" exploit a layer's expressive power during input-adaptive inference, enabling finer-grained accuracy-computational cost trade-offs. It presents a unified view to link input-adaptive layer skipping and input-adaptive hybrid quantization. Extensive experimental results demonstrate the superior tradeoff between computational cost and model expressive power (accuracy) achieved by DFS. More visualizations also indicate a smooth and consistent transition in the DFS behaviors, especially the learned choices between layer skipping and different quantizations when the total computational budgets vary, validating our hypothesis that layer quantization could be viewed as intermediate variants of layer skipping. Our source code and supplementary material are available at \link{https://github.com/Torment123/DFS}.
Quantization is a promising approach for reducing the inference time and memory footprint of neural networks. However, most existing quantization methods require access to the original training dataset for retraining during quantization. This is often not possible for applications with sensitive or proprietary data, e.g., due to privacy and security concerns. Existing zero-shot quantization methods use different heuristics to address this, but they result in poor performance, especially when quantizing to ultra-low precision. Here, we propose ZeroQ , a novel zero-shot quantization framework to address this. ZeroQ enables mixed-precision quantization without any access to the training or validation data. This is achieved by optimizing for a Distilled Dataset, which is engineered to match the statistics of batch normalization across different layers of the network. ZeroQ supports both uniform and mixed-precision quantization. For the latter, we introduce a novel Pareto frontier based method to automatically determine the mixed-precision bit setting for all layers, with no manual search involved. We extensively test our proposed method on a diverse set of models, including ResNet18/50/152, MobileNetV2, ShuffleNet, SqueezeNext, and InceptionV3 on ImageNet, as well as RetinaNet-ResNet50 on the Microsoft COCO dataset. In particular, we show that ZeroQ can achieve 1.71\% higher accuracy on MobileNetV2, as compared to the recently proposed DFQ method. Importantly, ZeroQ has a very low computational overhead, and it can finish the entire quantization process in less than 30s (0.5\% of one epoch training time of ResNet50 on ImageNet). We have open-sourced the ZeroQ framework\footnote{https://github.com/amirgholami/ZeroQ}.
Efficient model inference is an important and practical issue in the deployment of deep neural network on resource constraint platforms. Network quantization addresses this problem effectively by leveraging low-bit representation and arithmetic that could be conducted on dedicated embedded systems. In the previous works, the parameter bitwidth is set homogeneously and there is a trade-off between superior performance and aggressive compression. Actually the stacked network layers, which are generally regarded as hierarchical feature extractors, contribute diversely to the overall performance. For a well-trained neural network, the feature distributions of different categories differentiate gradually as the network propagates forward. Hence the capability requirement on the subsequent feature extractors is reduced. It indicates that the neurons in posterior layers could be assigned with lower bitwidth for quantized neural networks. Based on this observation, a simple but effective mixed-precision quantized neural network with progressively ecreasing bitwidth is proposed to improve the trade-off between accuracy and compression. Extensive experiments on typical network architectures and benchmark datasets demonstrate that the proposed method could achieve better or comparable results while reducing the memory space for quantized parameters by more than 30\% in comparison with the homogeneous counterparts. In addition, the results also demonstrate that the higher-precision bottom layers could boost the 1-bit network performance appreciably due to a better preservation of the original image information while the lower-precision posterior layers contribute to the regularization of $k-$bit networks.
Recently low-bit (e.g., 8-bit) network quantization has been extensively studied to accelerate the inference. Besides inference, low-bit training with quantized gradients can further bring more considerable acceleration, since the backward process is often computation-intensive. Unfortunately, the inappropriate quantization of backward propagation usually makes the training unstable and even crash. There lacks a successful unified low-bit training framework that can support diverse networks on various tasks. In this paper, we give an attempt to build a unified 8-bit (INT8) training framework for common convolutional neural networks from the aspects of both accuracy and speed. First, we empirically find the four distinctive characteristics of gradients, which provide us insightful clues for gradient quantization. Then, we theoretically give an in-depth analysis of the convergence bound and derive two principles for stable INT8 training. Finally, we propose two universal techniques, including Direction Sensitive Gradient Clipping that reduces the direction deviation of gradients and Deviation Counteractive Learning Rate Scaling that avoids illegal gradient update along the wrong direction. The experiments show that our unified solution promises accurate and efficient INT8 training for a variety of networks and tasks, including MobileNetV2, InceptionV3 and object detection that prior studies have never succeeded. Moreover, it enjoys a strong flexibility to run on off-the-shelf hardware, and reduces the training time by 22% on Pascal GPU without too much optimization effort. We believe that this pioneering study will help lead the community towards a fully unified INT8 training for convolutional neural networks.
Quantization reduces computation costs of neural networks but suffers from performance degeneration. Is this accuracy drop due to the reduced capacity, or inefficient training during the quantization procedure? After looking into the gradient propagation process of neural networks by viewing the weights and intermediate activations as random variables, we discover two critical rules for efficient training. Recent quantization approaches violates the two rules and results in degenerated convergence. To deal with this problem, we propose a simple yet effective technique, named scale-adjusted training (SAT), to comply with the discovered rules and facilitates efficient training. We also analyze the quantization error introduced in calculating the gradient in the popular parameterized clipping activation (PACT) technique. Through SAT together with gradient-calibrated PACT, quantized models obtain comparable or even better performance than their full-precision counterparts, achieving state-of-the-art accuracy with consistent improvement over previous quantization methods on a wide spectrum of models including MobileNet-V1/V2 and PreResNet-50.
The implementation of Deep Convolutional Neural Networks (ConvNets) on tiny end-nodes with limited non-volatile memory space calls for smart compression strategies capable of shrinking the footprint yet preserving predictive accuracy. There exist a number of strategies for this purpose, from those that play with the topology of the model or the arithmetic precision, e.g. pruning and quantization, to those that operate a model agnostic compression, e.g. weight encoding. The tighter the memory constraint, the higher the probability that these techniques alone cannot meet the requirement, hence more awareness and cooperation across different optimizations become mandatory. This work addresses the issue by introducing EAST, Encoding-Aware Sparse Training, a novel memory-constrained training procedure that leads quantized ConvNets towards deep memory compression. EAST implements an adaptive group pruning designed to maximize the compression rate of the weight encoding scheme (the LZ4 algorithm in this work). If compared to existing methods, EAST meets the memory constraint with lower sparsity, hence ensuring higher accuracy. Results conducted on a state-of-the-art ConvNet (ResNet-9) deployed on a low-power microcontroller (ARM Cortex-M4) validate the proposal.
Deep neural networks with adaptive configurations have gained increasing attention due to the instant and flexible deployment of these models on platforms with different resource budgets. In this paper, we investigate a novel option to achieve this goal by enabling adaptive bit-widths of weights and activations in the model. We first examine the benefits and challenges of training quantized model with adaptive bit-widths, and then experiment with several approaches including direct adaptation, progressive training and joint training. We discover that joint training is able to produce comparable performance on the adaptive model as individual models. We further propose a new technique named Switchable Clipping Level (S-CL) to further improve quantized models at the lowest bit-width. With our proposed techniques applied on a bunch of models including MobileNet-V1/V2 and ResNet-50, we demonstrate that bit-width of weights and activations is a new option for adaptively executable deep neural networks, offering a distinct opportunity for improved accuracy-efficiency trade-off as well as instant adaptation according to the platform constraints in real-world applications.
Quantization of weights of deep neural networks (DNN) has proven to be an effective solution for the purpose of implementing DNNs on edge devices such as mobiles, ASICs and FPGAs, because they have no sufficient resources to support computation involving millions of high precision weights and multiply-accumulate operations. This paper proposes a novel method to compress vectors of high precision weights of DNNs to ternary vectors, namely a cosine similarity based target non-retraining ternary (TNT) compression method. Our method leverages cosine similarity instead of Euclidean distances as commonly used in the literature and succeeds in reducing the size of the search space to find optimal ternary vectors from 3N to N, where N is the dimension of target vectors. As a result, the computational complexity for TNT to find theoretically optimal ternary vectors is only O(N log(N)). Moreover, our experiments show that, when we ternarize models of DNN with high precision parameters, the obtained quantized models can exhibit sufficiently high accuracy so that re-training models is not necessary.
Deep neural networks (DNNs) can be made hardware-efficient by reducing the numerical precision of the weights and activations of the network and by improving the network's resilience to noise. However, this gain in efficiency often comes at the cost of significantly reduced accuracy. In this paper, we present a novel approach to quantizing convolutional neural network. The resulting networks perform all computations in low-precision, without requiring higher-precision BN and nonlinearities, while still being highly accurate. To achieve this result, we employ a novel quantization technique that learns to optimally quantize the weights and activations of the network during training. Additionally, to enhance training convergence we use a new training technique, called gradual quantization. We leverage the nonlinear and normalizing behavior of our quantization function to effectively remove the higher-precision nonlinearities and BN from the network. The resulting convolutional layers are fully quantized to low precision, from input to output, ideal for neural network accelerators on the edge. We demonstrate the potential of this approach on different datasets and networks, showing that ternary-weight CNNs with low-precision in- and outputs perform virtually on par with their full-precision equivalents. Finally, we analyze the influence of noise on the weights, activations and convolution outputs (multiply-accumulate, MAC) and propose a strategy to improve network performance under noisy conditions.
We investigate the compression of deep neural networks by quantizing their weights and activations into multiple binary bases, known as multi-bit networks (MBNs), which accelerate the inference and reduce the storage for the deployment on low-resource mobile and embedded platforms. We propose Adaptive Loss-aware Quantization (ALQ), a new MBN quantization pipeline that is able to achieve an average bitwidth below one-bit without notable loss in inference accuracy. Unlike previous MBN quantization solutions that train a quantizer by minimizing the error to reconstruct full precision weights, ALQ directly minimizes the quantization-induced error on the loss function involving neither gradient approximation nor full precision maintenance. ALQ also exploits strategies including adaptive bitwidth, smooth bitwidth reduction, and iterative trained quantization to allow a smaller network size without loss in accuracy. Experiment results on popular image datasets show that ALQ outperforms state-of-the-art compressed networks in terms of both storage and accuracy. Code is available at https://github.com/zqu1992/ALQ
Similarity search retrieves the nearest neighbors of a query vector from a dataset of high-dimensional vectors. As the size of the dataset grows, the cost of performing the distance computations needed to implement a query can become prohibitive. A method often used to reduce this computational cost is quantization of the vector space and location-based encoding of the dataset vectors. These encodings can be used during query processing to find approximate nearest neighbors of the query point quickly. Search speed can be improved by using shorter codes, but shorter codes have higher quantization error, leading to degraded precision. In this work, we propose the Interleaved Composite Quantization (ICQ) which achieves fast similarity search without using shorter codes. In ICQ, a small subset of the code is used to approximate the distances, with complete codes being used only when necessary. Our method effectively reduces both code length and quantization error. Furthermore, ICQ is compatible with several recently proposed techniques for reducing quantization error and can be used in conjunction with these other techniques to improve results. We confirm these claims and show strong empirical performance of ICQ using several synthetic and real-word datasets.
Neural Networks are currently one of the most widely deployed machine learning algorithms. In particular, Convolutional Neural Networks (CNNs), are gaining popularity and are evaluated for deployment in safety critical applications such as self driving vehicles. Modern CNNs feature enormous memory bandwidth and high computational needs, challenging existing hardware platforms to meet throughput, latency and power requirements. Functional safety and error tolerance need to be considered as additional requirement in safety critical systems. In general, fault tolerant operation can be achieved by adding redundancy to the system, which is further exacerbating the computational demands. Furthermore, the question arises whether pruning and quantization methods for performance scaling turn out to be counterproductive with regards to fail safety requirements. In this work we present a methodology to evaluate the impact of permanent faults affecting Quantized Neural Networks (QNNs) and how to effectively decrease their effects in hardware accelerators. We use FPGA-based hardware accelerated error injection, in order to enable the fast evaluation. A detailed analysis is presented showing that QNNs containing convolutional layers are by far not as robust to faults as commonly believed and can lead to accuracy drops of up to 10%. To circumvent that, we propose two different methods to increase their robustness: 1) selective channel replication which adds significantly less redundancy than used by the common triple modular redundancy and 2) a fault-aware scheduling of processing elements for folded implementations
Recently it has been shown that deep learning-based image compression has shown the potential to outperform traditional codecs. However, most existing methods train multiple networks for multiple bit rates, which increases the implementation complexity. In this paper, we propose a variable-rate image compression framework, which employs more Generalized Divisive Normalization (GDN) layers than previous GDN-based methods. Novel GDN-based residual sub-networks are also developed in the encoder and decoder networks. Our scheme also uses a stochastic rounding-based scalable quantization. To further improve the performance, we encode the residual between the input and the reconstructed image from the decoder network as an enhancement layer. To enable a single model to operate with different bit rates and to learn multi-rate image features, a new objective function is introduced. Experimental results show that the proposed framework trained with variable-rate objective function outperforms all standard codecs such as H.265/HEVC-based BPG and state-of-the-art learning-based variable-rate methods.
Rocks under stress deform by creep mechanisms that include formation and slip on small-scale internal cracks. Intragranular cracks and slip along grain contacts release energy as elastic waves termed acoustic emissions (AE). AEs are thought to contain predictive information that can be used for fault failure forecasting. Here we present a method using unsupervised classification and an attention network to forecast labquakes using AE waveform features. Our data were generated in a laboratory setting using a biaxial shearing device with granular fault gouge intended to mimic the conditions of tectonic faults. Here we analyzed the temporal evolution of AEs generated throughout several hundred laboratory earthquake cycles. We used a Conscience Self-Organizing Map (CSOM) to perform topologically ordered vector quantization based on waveform properties. The resulting map was used to interactively cluster AEs. We examined the clusters over time to identify those with predictive ability. Finally, we used a variety of LSTM and attention-based networks to test the predictive power of the AE clusters. By tracking cumulative waveform features over the seismic cycle, the network is able to forecast the time-to-failure (TTF) of lab earthquakes. Our results show that analyzing the data to isolate predictive signals and using a more sophisticated network architecture are key to robustly forecasting labquakes. In the future, this method could be applied on tectonic faults monitor earthquakes and augment current early warning systems.
The computing wall and data movement challenges of deep neural networks (DNNs) have exposed the limitations of conventional CMOS-based DNN accelerators. Furthermore, the deep structure and large model size will make DNNs prohibitive to embedded systems and IoT devices, where low power consumption are required. To address these challenges, spin orbit torque magnetic random-access memory (SOT-MRAM) and SOT-MRAM based Processing-In-Memory (PIM) engines have been used to reduce the power consumption of DNNs since SOT-MRAM has the characteristic of near-zero standby power, high density, none-volatile. However, the drawbacks of SOT-MRAM based PIM engines such as high writing latency and requiring low bit-width data decrease its popularity as a favorable energy efficient DNN accelerator. To mitigate these drawbacks, we propose an ultra energy efficient framework by using model compression techniques including weight pruning and quantization from the software level considering the architecture of SOT-MRAM PIM. And we incorporate the alternating direction method of multipliers (ADMM) into the training phase to further guarantee the solution feasibility and satisfy SOT-MRAM hardware constraints. Thus, the footprint and power consumption of SOT-MRAM PIM can be reduced, while increasing the overall system throughput at the meantime, making our proposed ADMM-based SOT-MRAM PIM more energy efficiency and suitable for embedded systems or IoT devices. Our experimental results show the accuracy and compression rate of our proposed framework is consistently outperforming the reference works, while the efficiency (area \& power) and throughput of SOT-MRAM PIM engine is significantly improved.
Three dimensional convolutional neural networks (3DCNNs) have been applied in many tasks, e.g., video and 3D point cloud recognition. However, due to the higher dimension of convolutional kernels, the space complexity of 3DCNNs is generally larger than that of traditional two dimensional convolutional neural networks (2DCNNs). To miniaturize 3DCNNs for the deployment in confining environments such as embedded devices, neural network compression is a promising approach. In this work, we adopt the tensor train (TT) decomposition, a straightforward and simple in situ training compression method, to shrink the 3DCNN models. Through proposing tensorizing 3D convolutional kernels in TT format, we investigate how to select appropriate TT ranks for achieving higher compression ratio. We have also discussed the redundancy of 3D convolutional kernels for compression, core significance and future directions of this work, as well as the theoretical computation complexity versus practical executing time of convolution in TT. In the light of multiple contrast experiments based on VIVA challenge, UCF11, and UCF101 datasets, we conclude that TT decomposition can compress 3DCNNs by around one hundred times without significant accuracy loss, which will enable its applications in extensive real world scenarios.
Higher-order tensors can represent scores in a rating system, frames in a video, and images of the same subject. In practice, the measurements are often highly quantized due to the sampling strategies or the quality of devices. Existing works on tensor recovery have focused on data losses and random noises. Only a few works consider tensor recovery from quantized measurements but are restricted to binary measurements. This paper, for the first time, addresses the problem of tensor recovery from multi-level quantized measurements. Leveraging the low-rank property of the tensor, this paper proposes a nonconvex optimization problem for tensor recovery. We provide a theoretical upper bound of the recovery error, which diminishes to zero when the sizes of dimensions increase to infinity. Our error bound significantly improves over the existing results in one-bit tensor recovery and quantized matrix recovery. A tensor-based alternating proximal gradient descent algorithm with a convergence guarantee is proposed to solve the nonconvex problem. Our recovery method can handle data losses and do not need the information of the quantization rule. The method is validated on synthetic data, image datasets, and music recommender datasets.
Besides accuracy, the model size of convolutional neural networks (CNN) models is another important factor considering limited hardware resources in practical applications. For example, employing deep neural networks on mobile systems requires the design of accurate yet fast CNN for low latency in classification and object detection. To fulfill the need, we aim at obtaining CNN models with both high testing accuracy and small size to address resource constraints in many embedded devices. In particular, this paper focuses on proposing a generic reinforcement learning-based model compression approach in a two-stage compression pipeline: pruning and quantization. The first stage of compression, i.e., pruning, is achieved via exploiting deep reinforcement learning (DRL) to co-learn the accuracy and the FLOPs updated after layer-wise channel pruning and element-wise variational pruning via information dropout. The second stage, i.e., quantization, is achieved via a similar DRL approach but focuses on obtaining the optimal bits representation for individual layers. We further conduct experimental results on CIFAR-10 and ImageNet datasets. For the CIFAR-10 dataset, the proposed method can reduce the size of VGGNet by 9x from 20.04MB to 2.2MB with a slight accuracy increase. For the ImageNet dataset, the proposed method can reduce the size of VGG-16 by 33x from 138MB to 4.14MB with no accuracy loss.