Keyword spotting has gained popularity as a natural way to interact with consumer devices in recent years. However, because of its always-on nature and the variety of speech, it necessitates a low-power design as well as user customization. This paper describes a low-power, energy-efficient keyword spotting accelerator with SRAM based in-memory computing (IMC) and on-chip learning for user customization. However, IMC is constrained by macro size, limited precision, and non-ideal effects. To address the issues mentioned above, this paper proposes bias compensation and fine-tuning using an IMC-aware model design. Furthermore, because learning with low-precision edge devices results in zero error and gradient values due to quantization, this paper proposes error scaling and small gradient accumulation to achieve the same accuracy as ideal model training. The simulation results show that with user customization, we can recover the accuracy loss from 51.08\% to 89.76\% with compensation and fine-tuning and further improve to 96.71\% with customization. The chip implementation can successfully run the model with only 14$uJ$ per decision. When compared to the state-of-the-art works, the presented design has higher energy efficiency with additional on-chip model customization capabilities for higher accuracy.
Transformers have achieved great success in pluralistic image inpainting recently. However, we find existing transformer based solutions regard each pixel as a token, thus suffer from information loss issue from two aspects: 1) They downsample the input image into much lower resolutions for efficiency consideration, incurring information loss and extra misalignment for the boundaries of masked regions. 2) They quantize $256^3$ RGB pixels to a small number (such as 512) of quantized pixels. The indices of quantized pixels are used as tokens for the inputs and prediction targets of transformer. Although an extra CNN network is used to upsample and refine the low-resolution results, it is difficult to retrieve the lost information back.To keep input information as much as possible, we propose a new transformer based framework "PUT". Specifically, to avoid input downsampling while maintaining the computation efficiency, we design a patch-based auto-encoder P-VQVAE, where the encoder converts the masked image into non-overlapped patch tokens and the decoder recovers the masked regions from inpainted tokens while keeping the unmasked regions unchanged. To eliminate the information loss caused by quantization, an Un-Quantized Transformer (UQ-Transformer) is applied, which directly takes the features from P-VQVAE encoder as input without quantization and regards the quantized tokens only as prediction targets. Extensive experiments show that PUT greatly outperforms state-of-the-art methods on image fidelity, especially for large masked regions and complex large-scale datasets. Code is available at https://github.com/liuqk3/PUT
This paper addresses the challenges of training large neural network models under federated learning settings: high on-device memory usage and communication cost. The proposed Online Model Compression (OMC) provides a framework that stores model parameters in a compressed format and decompresses them only when needed. We use quantization as the compression method in this paper and propose three methods, (1) using per-variable transformation, (2) weight matrices only quantization, and (3) partial parameter quantization, to minimize the impact on model accuracy. According to our experiments on two recent neural networks for speech recognition and two different datasets, OMC can reduce memory usage and communication cost of model parameters by up to 59% while attaining comparable accuracy and training speed when compared with full-precision training.
This paper tackles the problem of ensuring training data privacy in a federated learning context. Relying on Homomorphic Encryption (HE) and Differential Privacy (DP), we propose a framework addressing threats on the privacy of the training data. Notably, the proposed framework ensures the privacy of the training data from all actors of the learning process, namely the data owners and the aggregating server. More precisely, while HE blinds a semi-honest server during the learning protocol, DP protects the data from semi-honest clients participating in the training process as well as end-users with black-box or white-box access to the trained model. In order to achieve this, we provide new theoretical and practical results to allow these techniques to be rigorously combined. In particular, by means of a novel stochastic quantisation operator, we prove DP guarantees in a context where the noise is quantised and bounded due to the use of HE. The paper is concluded by experiments which show the practicality of the entire framework in terms of both model quality (impacted by DP) and computational overhead (impacted by HE).
We consider the image and video compression on resource limited platforms. An ultra low-cost image encoder, named Block Modulating Video Compression (BMVC) with an encoding complexity ${\cal O}(1)$ is proposed to be implemented on mobile platforms with low consumption of power and computation resources. We also develop two types of BMVC decoders, implemented by deep neural networks. The first BMVC decoder is based on the Plug-and-Play (PnP) algorithm, which is flexible to different compression ratios. And the second decoder is a memory efficient end-to-end convolutional neural network, which aims for real-time decoding. Extensive results on the high definition images and videos demonstrate the superior performance of the proposed codec and the robustness against bit quantization.
Federated learning is a machine learning training paradigm that enables clients to jointly train models without sharing their own localized data. However, the implementation of federated learning in practice still faces numerous challenges, such as the large communication overhead due to the repetitive server-client synchronization and the lack of adaptivity by SGD-based model updates. Despite that various methods have been proposed for reducing the communication cost by gradient compression or quantization, and the federated versions of adaptive optimizers such as FedAdam are proposed to add more adaptivity, the current federated learning framework still cannot solve the aforementioned challenges all at once. In this paper, we propose a novel communication-efficient adaptive federated learning method (FedCAMS) with theoretical convergence guarantees. We show that in the nonconvex stochastic optimization setting, our proposed FedCAMS achieves the same convergence rate of $O(\frac{1}{\sqrt{TKm}})$ as its non-compressed counterparts. Extensive experiments on various benchmarks verify our theoretical analysis.
Memristors enable the computation of matrix-vector multiplications (MVM) in memory and, therefore, show great potential in highly increasing the energy efficiency of deep neural network (DNN) inference accelerators. However, computations in memristors suffer from hardware non-idealities and are subject to different sources of noise that may negatively impact system performance. In this work, we theoretically analyze the mean squared error of DNNs that use memristor crossbars to compute MVM. We take into account both the quantization noise, due to the necessity of reducing the DNN model size, and the programming noise, stemming from the variability during the programming of the memristance value. Simulations on pre-trained DNN models showcase the accuracy of the analytical prediction. Furthermore the proposed method is almost two order of magnitude faster than Monte-Carlo simulation, thus making it possible to optimize the implementation parameters to achieve minimal error for a given power constraint.
To obtain lower inference latency and less memory footprint of deep neural networks, model quantization has been widely employed in deep model deployment, by converting the floating points to low-precision integers. However, previous methods (such as quantization aware training and post training quantization) require original data for the fine-tuning or calibration of quantized model, which makes them inapplicable to the cases that original data are not accessed due to privacy or security. This gives birth to the data-free quantization method with synthetic data generation. While current data-free quantization methods still suffer from severe performance degradation when quantizing a model into lower bit, caused by the low inter-class separability of semantic features. To this end, we propose a new and effective data-free quantization method termed ClusterQ, which utilizes the feature distribution alignment for synthetic data generation. To obtain high inter-class separability of semantic features, we cluster and align the feature distribution statistics to imitate the distribution of real data, so that the performance degradation is alleviated. Moreover, we incorporate the diversity enhancement to solve class-wise mode collapse. We also employ the exponential moving average to update the centroid of each cluster for further feature distribution improvement. Extensive experiments based on different deep models (e.g., ResNet-18 and MobileNet-V2) over the ImageNet dataset demonstrate that our proposed ClusterQ model obtains state-of-the-art performance.
Point cloud compression plays a crucial role in reducing the huge cost of data storage and transmission. However, distortions can be introduced into the decompressed point clouds due to quantization. In this paper, we propose a novel learning-based post-processing method to enhance the decompressed point clouds. Specifically, a voxelized point cloud is first divided into small cubes. Then, a 3D convolutional network is proposed to predict the occupancy probability for each location of a cube. We leverage both local and global contexts by generating multi-scale probabilities. These probabilities are progressively summed to predict the results in a coarse-to-fine manner. Finally, we obtain the geometry-refined point clouds based on the predicted probabilities. Different from previous methods, we deal with decompressed point clouds with huge variety of distortions using a single model. Experimental results show that the proposed method can significantly improve the quality of the decompressed point clouds, achieving 9.30dB BDPSNR gain on three representative datasets on average.
We introduce a Power-of-Two low-bit post-training quantization(PTQ) method for deep neural network that meets hardware requirements and does not call for long-time retraining. Power-of-Two quantization can convert the multiplication introduced by quantization and dequantization to bit-shift that is adopted by many efficient accelerators. However, the Power-of-Two scale factors have fewer candidate values, which leads to more rounding or clipping errors. We propose a novel Power-of-Two PTQ framework, dubbed RAPQ, which dynamically adjusts the Power-of-Two scales of the whole network instead of statically determining them layer by layer. It can theoretically trade off the rounding error and clipping error of the whole network. Meanwhile, the reconstruction method in RAPQ is based on the BN information of every unit. Extensive experiments on ImageNet prove the excellent performance of our proposed method. Without bells and whistles, RAPQ can reach accuracy of 65% and 48% on ResNet-18 and MobileNetV2 respectively with weight INT2 activation INT4. We are the first to propose the more constrained but hardware-friendly Power-of-Two quantization scheme for low-bit PTQ specially and prove that it can achieve nearly the same accuracy as SOTA PTQ method. The code was released.
Recently, federated learning (FL), which replaces data sharing with model sharing, has emerged as an efficient and privacy-friendly machine learning (ML) paradigm. One of the main challenges in FL is the huge communication cost for model aggregation. Many compression/quantization schemes have been proposed to reduce the communication cost for model aggregation. However, the following question remains unanswered: What is the fundamental trade-off between the communication cost and the FL convergence performance? In this paper, we manage to answer this question. Specifically, we first put forth a general framework for model aggregation performance analysis based on the rate-distortion theory. Under the proposed analysis framework, we derive an inner bound of the rate-distortion region of model aggregation. We then conduct an FL convergence analysis to connect the aggregation distortion and the FL convergence performance. We formulate an aggregation distortion minimization problem to improve the FL convergence performance. Two algorithms are developed to solve the above problem. Numerical results on aggregation distortion, convergence performance, and communication cost demonstrate that the baseline model aggregation schemes still have great potential for further improvement.
Deep neural networks (DNNs) have achieved unprecedented success in the field of artificial intelligence (AI), including computer vision, natural language processing and speech recognition. However, their superior performance comes at the considerable cost of computational complexity, which greatly hinders their applications in many resource-constrained devices, such as mobile phones and Internet of Things (IoT) devices. Therefore, methods and techniques that are able to lift the efficiency bottleneck while preserving the high accuracy of DNNs are in great demand in order to enable numerous edge AI applications. This paper provides an overview of efficient deep learning methods, systems and applications. We start from introducing popular model compression methods, including pruning, factorization, quantization as well as compact model design. To reduce the large design cost of these manual solutions, we discuss the AutoML framework for each of them, such as neural architecture search (NAS) and automated pruning and quantization. We then cover efficient on-device training to enable user customization based on the local data on mobile devices. Apart from general acceleration techniques, we also showcase several task-specific accelerations for point cloud, video and natural language processing by exploiting their spatial sparsity and temporal/token redundancy. Finally, to support all these algorithmic advancements, we introduce the efficient deep learning system design from both software and hardware perspectives.
MOS (Mean Opinion Score) is a subjective method used for the evaluation of a system's quality. Telecommunications (for voice and video), and speech synthesis systems (for generated speech) are a few of the many applications of the method. While MOS tests are widely accepted, they are time-consuming and costly since human input is required. In addition, since the systems and subjects of the tests differ, the results are not really comparable. On the other hand, a large number of previous tests allow us to train machine learning models that are capable of predicting MOS value. By automatically predicting MOS values, both the aforementioned issues can be resolved. The present work introduces data-, training- and post-training specific improvements to a previous self-supervised learning-based MOS prediction model. We used a wav2vec 2.0 model pre-trained on LibriSpeech, extended with LSTM and non-linear dense layers. We introduced transfer learning, target data preprocessing a two- and three-phase training method with different batch formulations, dropout accumulation (for larger batch sizes) and quantization of the predictions. The methods are evaluated using the shared synthetic speech dataset of the first Voice MOS challenge.
Full-precision deep learning models are typically too large or costly to deploy on edge devices. To accommodate to the limited hardware resources, models are adapted to the edge using various edge-adaptation techniques, such as quantization and pruning. While such techniques may have a negligible impact on top-line accuracy, the adapted models exhibit subtle differences in output compared to the original model from which they are derived. In this paper, we introduce a new evasive attack, DIVA, that exploits these differences in edge adaptation, by adding adversarial noise to input data that maximizes the output difference between the original and adapted model. Such an attack is particularly dangerous, because the malicious input will trick the adapted model running on the edge, but will be virtually undetectable by the original model, which typically serves as the authoritative model version, used for validation, debugging and retraining. We compare DIVA to a state-of-the-art attack, PGD, and show that DIVA is only 1.7-3.6% worse on attacking the adapted model but 1.9-4.2 times more likely not to be detected by the the original model under a whitebox and semi-blackbox setting, compared to PGD.
Recently, many deep learning based researches are conducted to explore the potential quality improvement of compressed videos. These methods mostly utilize either the spatial or temporal information to perform frame-level video enhancement. However, they fail in combining different spatial-temporal information to adaptively utilize adjacent patches to enhance the current patch and achieve limited enhancement performance especially on scene-changing and strong-motion videos. To overcome these limitations, we propose a patch-wise spatial-temporal quality enhancement network which firstly extracts spatial and temporal features, then recalibrates and fuses the obtained spatial and temporal features. Specifically, we design a temporal and spatial-wise attention-based feature distillation structure to adaptively utilize the adjacent patches for distilling patch-wise temporal features. For adaptively enhancing different patch with spatial and temporal information, a channel and spatial-wise attention fusion block is proposed to achieve patch-wise recalibration and fusion of spatial and temporal features. Experimental results demonstrate our network achieves peak signal-to-noise ratio improvement, 0.55 – 0.69 dB compared with the compressed videos at different quantization parameters, outperforming state- of-the-art approaches.
Most studies in cross-device federated learning focus on small models, due to the server-client communication and on-device computation bottlenecks. In this work, we leverage various techniques for mitigating these bottlenecks to train larger language models in cross-device federated learning. With systematic applications of partial model training, quantization, efficient transfer learning, and communication-efficient optimizers, we are able to train a $21$M parameter Transformer and $20.2$M parameter Conformer that achieve the same or better perplexity as that of a similarly sized LSTM with $\sim10\times$ smaller client-to-server communication cost and $11\%$ lower perplexity than smaller LSTMs commonly studied in literature.
Conventional model quantization methods use a fixed quantization scheme to different data samples, which ignores the inherent "recognition difficulty" differences between various samples. We propose to feed different data samples with varying quantization schemes to achieve a data-dependent dynamic inference, at a fine-grained layer level. However, enabling this adaptive inference with changeable layer-wise quantization schemes is challenging because the combination of bit-widths and layers is growing exponentially, making it extremely difficult to train a single model in such a vast searching space and use it in practice. To solve this problem, we present the Arbitrary Bit-width Network (ABN), where the bit-widths of a single deep network can change at runtime for different data samples, with a layer-wise granularity. Specifically, first we build a weight-shared layer-wise quantizable "super-network" in which each layer can be allocated with multiple bit-widths and thus quantized differently on demand. The super-network provides a considerably large number of combinations of bit-widths and layers, each of which can be used during inference without retraining or storing myriad models. Second, based on the well-trained super-network, each layer's runtime bit-width selection decision is modeled as a Markov Decision Process (MDP) and solved by an adaptive inference strategy accordingly. Experiments show that the super-network can be built without accuracy degradation, and the bit-widths allocation of each layer can be adjusted to deal with various inputs on the fly. On ImageNet classification, we achieve 1.1% top1 accuracy improvement while saving 36.2% BitOps.
This paper introduces CO3 -- an algorithm for communication-efficient federated Deep Neural Network (DNN) training. CO3 takes its name from three processing applied which reduce the communication load when transmitting the local DNN gradients from the remote users to the Parameter Server. Namely: (i) gradient quantization through floating-point conversion, (ii) lossless compression of the quantized gradient, and (iii) quantization error correction. We carefully design each of the steps above to assure good training performance under a constraint on the communication rate. In particular, in steps (i) and (ii), we adopt the assumption that DNN gradients are distributed according to a generalized normal distribution, which is validated numerically in the paper. For step (iii), we utilize an error feedback with memory decay mechanism to correct the quantization error introduced in step (i). We argue that the memory decay coefficient, similarly to the learning rate, can be optimally tuned to improve convergence. A rigorous convergence analysis of the proposed CO3 with SGD is provided. Moreover, with extensive simulations, we show that CO3 offers improved performance when compared with existing gradient compression schemes in the literature which employ sketching and non-uniform quantization of the local gradients.
Although deep generative models have gained a lot of attention, most of the existing works are designed for unimodal generation. In this paper, we explore a new method for unconditional image-text pair generation. We design Multimodal Cross-Quantization VAE (MXQ-VAE), a novel vector quantizer for joint image-text representations, with which we discover that a joint image-text representation space is effective for semantically consistent image-text pair generation. To learn a multimodal semantic correlation in a quantized space, we combine VQ-VAE with a Transformer encoder and apply an input masking strategy. Specifically, MXQ-VAE accepts a masked image-text pair as input and learns a quantized joint representation space, so that the input can be converted to a unified code sequence, then we perform unconditional image-text pair generation with the code sequence. Extensive experiments show the correlation between the quantized joint space and the multimodal generation capability on synthetic and real-world datasets. In addition, we demonstrate the superiority of our approach in these two aspects over several baselines. The source code is publicly available at: https://github.com/ttumyche/MXQ-VAE.
Binary Neural Networks (BNNs) have emerged as a promising solution for reducing the memory footprint and compute costs of deep neural networks, but they suffer from quality degradation due to the lack of freedom as activations and weights are constrained to the binary values. To compensate for the accuracy drop, we propose a novel BNN design called Binary Neural Network with INSTAnce-aware threshold (INSTA-BNN), which controls the quantization threshold dynamically in an input-dependent or instance-aware manner. According to our observation, higher-order statistics can be a representative metric to estimate the characteristics of the input distribution. INSTA-BNN is designed to adjust the threshold dynamically considering various information, including higher-order statistics, but it is also optimized judiciously to realize minimal overhead on a real device. Our extensive study shows that INSTA-BNN outperforms the baseline by 3.0% and 2.8% on the ImageNet classification task with comparable computing cost, achieving 68.5% and 72.2% top-1 accuracy on ResNet-18 and MobileNetV1 based models, respectively.