We demonstrate an image dequantizing diffusion model that enables novel edits on natural images. We propose operating on quantized images because they offer easy abstraction for patch-based edits and palette transfer. In particular, we show that color palettes can make the output of the diffusion model easier to control and interpret. We first establish that existing image restoration methods are not sufficient, such as JPEG noise reduction models. We then demonstrate that our model can generate natural images that respect the color palette the user asked for. For palette transfer, we propose a method based on weighted bipartite matching. We then show that our model generates plausible images even after extreme palette transfers, respecting user query. Our method can optionally condition on the source texture in part or all of the image. In doing so, we overcome a common problem in existing image colorization methods that are unable to produce colors with a different luminance than the input. We evaluate several possibilities for texture conditioning and their trade-offs, including luminance, image gradients, and thresholded gradients, the latter of which performed best in maintaining texture and color control simultaneously. Our method can be usefully extended to another practical edit: recoloring patches of an image while respecting the source texture. Our procedure is supported by several qualitative and quantitative evaluations.
We study efficient mechanisms for differentially private kernel density estimation (DP-KDE). Prior work for the Gaussian kernel described algorithms that run in time exponential in the number of dimensions $d$. This paper breaks the exponential barrier, and shows how the KDE can privately be approximated in time linear in $d$, making it feasible for high-dimensional data. We also present improved bounds for low-dimensional data. Our results are obtained through a general framework, which we term Locality Sensitive Quantization (LSQ), for constructing private KDE mechanisms where existing KDE approximation techniques can be applied. It lets us leverage several efficient non-private KDE methods -- like Random Fourier Features, the Fast Gauss Transform, and Locality Sensitive Hashing -- and ``privatize'' them in a black-box manner. Our experiments demonstrate that our resulting DP-KDE mechanisms are fast and accurate on large datasets in both high and low dimensions.
Simultaneous operation of all sensors in a large-scale sensor network is power-consuming and computationally expensive. Hence, it is desirable to select fewer sensors. A greedy algorithm is widely used for sensor selection in homogeneous networks with a theoretical worst-case performance of (1-1/e) ~ 63% of the optimal performance when optimizing submodular metrics. For heterogeneous sensor networks (HSNs) comprising multiple sets of sensors, most of the existing sensor selection methods optimize the performance constrained by a budget on the total value of the selected sensors. However, in many applications, the number of sensors to select from each set is known apriori, and solutions are not well-explored. For this problem, we propose a joint greedy heterogeneous sensor selection algorithm. Theoretically, we show that the worst-case performance of the proposed algorithm is bounded to 50% of the optimum for submodular cost metrics. In the special case of HSNs with two sensor networks, the performance guarantee can be improved to 63% when the number of sensors to select from one set is much smaller than the other. To validate our results experimentally, we propose a submodular metric based on the frame potential measure that considers both the correlation among the sensor measurements and their heterogeneity. We prove theoretical bounds for the mean squared error of the solution when this performance metric is used. We validate our results through simulation experiments considering both linear and non-linear measurement models corrupted by additive noise and quantization errors. Our experiments show that the proposed algorithm results in 4-10 dB lower error than existing methods.
As radar systems will be equipped with thousands of antenna elements and wide bandwidth, the associated costs and power consumption become exceedingly high, and a potential solution is to adopt low-resolution quantization technology, which not only reduces data storage needs but also lowers power and hardware costs. This paper focuses on line spectral estimation and detection (LSE\&D) with few-bit ADCs (typically 1-4 bits) by investigating the signal-to-noise ratio (SNR) loss, establishing a framework to understand the impact of intersinusoidal interference, the bit-depth of the quantizer, and the noise variance on weak signal detection in scenarios involving multiple sinusoids under low-resolution quantization. Additionally, a low-complexity, super-resolution, and constant false alarm rate (CFAR) algorithm, named generalized Newtonized orthogonal matching pursuit (GNOMP), is proposed. Extensive numerical simulations are conducted to validate the theoretical findings, particularly in terms of the detection probability bound. The effectiveness of GNOMP is demonstrated through comparisons with state-of-the-art algorithms, the Cram\'{e}r Rao bound, and the detection probability bound. Real data acquired by mmWave radar further substantiates the effectiveness of GNOMP in practical applications.
Low-resolution analog-to-digital converters (ADCs) have emerged as an efficient solution for massive multiple-input multiple-output (MIMO) systems to reap high data rates with reasonable power consumption and hardware complexity. In this paper, we analyze the performance of oversampling in uplink massive MIMO orthogonal frequency-division multiplexing (MIMO-OFDM) systems with low-resolution ADCs. Considering both the temporal and spatial correlation of the quantization distortion, we derive an approximate closed-form expression of an achievable sum rate, which reveals how the oversampling ratio (OSR), the ADC resolution, and the signal-to-noise ratio (SNR) jointly affect the system performance. In particular, we demonstrate that oversampling can effectively improve the sum rate by mitigating the impact of the quantization distortion, especially at high SNR and with very low ADC resolution. Furthermore, we show that the considered low-resolution massive MIMO-OFDM system can achieve the same performance as the unquantized one when both the SNR and the OSR are sufficiently high. Numerical simulations confirm our analysis.
In this paper, the trade-off between the quantization noise and the dynamic range of ADCs used to acquire radar signals is revisited using the Unlimited Sensing Framework (USF) in a practical setting. Trade-offs between saturation and resolution arise in many applications, like radar, where sensors acquire signals which exhibit a high degree of variability in amplitude. To solve this issue, we propose the use of the co-design approach of the USF which acquires folded version of the signal of interest and leverages its structure to reconstruct it after its acquisition. We demonstrate that this method outperforms other standard acquisition methods for Doppler radars. Taking our theory all the way to practice, we develop a prototype USF-enabled Doppler Radar and show the clear benefits of our method. In each experiment, we show that using the USF increases sensitivity compared to a classic acquisition approach.
A growing number of applications depend on Machine Learning (ML) functionality and benefits from both higher quality ML predictions and better timeliness (latency) at the same time. A growing body of research in computer architecture, ML, and systems software literature focuses on reaching better latency-accuracy tradeoffs for ML models. Efforts include compression, quantization, pruning, early-exit models, mixed DNN precision, as well as ML inference accelerator designs that minimize latency and energy, while preserving delivered accuracy. All of them, however, yield improvements for a single static point in the latency-accuracy tradeoff space. We make a case for applications that operate in dynamically changing deployment scenarios, where no single static point is optimal. We draw on a recently proposed weight-shared SuperNet mechanism to enable serving a stream of queries that uses (activates) different SubNets within this weight-shared construct. This creates an opportunity to exploit the inherent temporal locality with our proposed SubGraph Stationary (SGS) optimization. We take a hardware-software co-design approach with a real implementation of SGS in SushiAccel and the implementation of a software scheduler SushiSched controlling which SubNets to serve and what to cache in real-time. Combined, they are vertically integrated into SUSHI-an inference serving stack. For the stream of queries, SUSHI yields up to 25% improvement in latency, 0.98% increase in served accuracy. SUSHI can achieve up to 78.7% off-chip energy savings.
A ReLU neural network leads to a finite polyhedral decomposition of input space and a corresponding finite dual graph. We show that while this dual graph is a coarse quantization of input space, it is sufficiently robust that it can be combined with persistent homology to detect homological signals of manifolds in the input space from samples. This property holds for a variety of networks trained for a wide range of purposes that have nothing to do with this topological application. We found this feature to be surprising and interesting; we hope it will also be useful.
We discuss a relationship between rate-distortion and optimal transport (OT) theory, even though they seem to be unrelated at first glance. In particular, we show that a function defined via an extremal entropic OT distance is equivalent to the rate-distortion function. We numerically verify this result as well as previous results that connect the Monge and Kantorovich problems to optimal scalar quantization. Thus, we unify solving scalar quantization and rate-distortion functions in an alternative fashion by using their respective optimal transport solvers.
Neural network quantization is a very promising solution in the field of model compression, but its resulting accuracy highly depends on a training/fine-tuning process and requires the original data. This not only brings heavy computation and time costs but also is not conducive to privacy and sensitive information protection. Therefore, a few recent works are starting to focus on data-free quantization. However, data-free quantization does not perform well while dealing with ultra-low precision quantization. Although researchers utilize generative methods of synthetic data to address this problem partially, data synthesis needs to take a lot of computation and time. In this paper, we propose a data-free mixed-precision compensation (DF-MPC) method to recover the performance of an ultra-low precision quantized model without any data and fine-tuning process. By assuming the quantized error caused by a low-precision quantized layer can be restored via the reconstruction of a high-precision quantized layer, we mathematically formulate the reconstruction loss between the pre-trained full-precision model and its layer-wise mixed-precision quantized model. Based on our formulation, we theoretically deduce the closed-form solution by minimizing the reconstruction loss of the feature maps. Since DF-MPC does not require any original/synthetic data, it is a more efficient method to approximate the full-precision model. Experimentally, our DF-MPC is able to achieve higher accuracy for an ultra-low precision quantized model compared to the recent methods without any data and fine-tuning process.
Deep neural networks (DNNs) offer the highest performance in a wide range of applications in computer vision. These results rely on over-parameterized backbones, which are expensive to run. This computational burden can be dramatically reduced by quantizing (in either data-free (DFQ), post-training (PTQ) or quantization-aware training (QAT) scenarios) floating point values to ternary values (2 bits, with each weight taking value in {-1,0,1}). In this context, we observe that rounding to nearest minimizes the expected error given a uniform distribution and thus does not account for the skewness and kurtosis of the weight distribution, which strongly affects ternary quantization performance. This raises the following question: shall one minimize the highest or average quantization error? To answer this, we design two operators: TQuant and MQuant that correspond to these respective minimization tasks. We show experimentally that our approach allows to significantly improve the performance of ternary quantization through a variety of scenarios in DFQ, PTQ and QAT and give strong insights to pave the way for future research in deep neural network quantization.
Despite the outstanding performance of transformers in both language and vision tasks, the expanding computation and model size have increased the demand for efficient deployment. To address the heavy computation and parameter drawbacks, quantization is frequently studied in the community as a representative model compression technique and has seen extensive use on ConvNets. However, due to the unique properties of transformers, the low-bit quantization applications are still limited and underexplored. In this paper, we identify the difficulty of transformer low-bit quantization-aware training on its unique variation behaviors, which significantly differ from ConvNets. Based on comprehensive quantitative analysis, we observe variation in three hierarchies: various module quantization sensitivities, outliers in static weight and activation distribution, and oscillation in dynamic parameter fluctuations. These variations of transformers bring instability to the quantization-aware training (QAT) and negatively influence the performance. We explore the best practices to alleviate the variation's influence during low-bit transformer QAT and propose a variation-aware quantization scheme for both vision and language transformers. We extensively verify and show our scheme can alleviate the variation and improve the performance of transformers across various models and tasks. Our solution substantially improves the 2-bit Swin-T and binary BERT-base, achieving a 3.35% and 1.4% accuracy improvement over previous state-of-the-art methods on ImageNet-1K and GLUE. Codes and models are available at https://github.com/HuangOwen/Quantization-Variation.
This work is devoted to the study of the influence of quantization noise on the spectral characteristics of a digital signal and the assessment of spectrum measurement errors that arise due to the quantization noise of an analog-to-digital converter. To achieve more accurate and reliable measurements of the spectrum, an error assessment was carried out, which allows taking into account the impact of quantization noise on the spectral data. This is important for obtaining more accurate results and ensuring high-quality measurements of the spectral components of the vibration signal. In addition, further research is aimed at developing methods for estimating spectrum measurement errors taking into account other possible sources of errors and contributing to the development of compensation algorithms to reduce the impact of quantization noise.
We propose a generalized convolutional neural network (CNN) architecture that first decomposes the input signal into subbands by an adaptive filter bank structure, and then uses convolutional layers to extract features from each subband independently. Fully connected layers finally combine the extracted features to perform classification. The proposed architecture restrains each of the subband CNNs from learning using the entire input signal spectrum, resulting in structural regularization. Our proposed CNN architecture is fully compatible with the end-to-end learning mechanism of typical CNN architectures and learns the subband decomposition from the input dataset. We show that the proposed CNN architecture has attractive properties, such as robustness to input and weight-and-bias quantization noise, compared to regular full-band CNN architectures. Importantly, the proposed architecture significantly reduces computational costs, while maintaining state-of-the-art classification accuracy. Experiments on image classification tasks using the MNIST, CIFAR-10/100, Caltech-101, and ImageNet-2012 datasets show that the proposed architecture allows accuracy surpassing state-of-the-art results. On the ImageNet-2012 dataset, we achieved top-5 and top-1 validation set accuracy of 86.91% and 69.73%, respectively. Notably, the proposed architecture offers over 90% reduction in computation cost in the inference path and approximately 75% reduction in back-propagation (per iteration) with just a single-layer subband decomposition. With a 2-layer subband decomposition, the computational gains are even more significant with comparable accuracy results to the single-layer decomposition.
Quantization is commonly used in Deep Neural Networks (DNNs) to reduce the storage and computational complexity by decreasing the arithmetical precision of activations and weights, a.k.a. tensors. Efficient hardware architectures employ linear quantization to enable the deployment of recent DNNs onto embedded systems and mobile devices. However, linear uniform quantization cannot usually reduce the numerical precision to less than 8 bits without sacrificing high performance in terms of model accuracy. The performance loss is due to the fact that tensors do not follow uniform distributions. In this paper, we show that a significant amount of tensors fit into an exponential distribution. Then, we propose DNA-TEQ to exponentially quantize DNN tensors with an adaptive scheme that achieves the best trade-off between numerical precision and accuracy loss. The experimental results show that DNA-TEQ provides a much lower quantization bit-width compared to previous proposals, resulting in an average compression ratio of 40% over the linear INT8 baseline, with negligible accuracy loss and without retraining the DNNs. Besides, DNA-TEQ leads the way in performing dot-product operations in the exponential domain, which saves 66% of energy consumption on average for a set of widely used DNNs.
Implicit Neural Representation (INR) is an innovative approach for representing complex shapes or objects without explicitly defining their geometry or surface structure. Instead, INR represents objects as continuous functions. Previous research has demonstrated the effectiveness of using neural networks as INR for image compression, showcasing comparable performance to traditional methods such as JPEG. However, INR holds potential for various applications beyond image compression. This paper introduces Rapid-INR, a novel approach that utilizes INR for encoding and compressing images, thereby accelerating neural network training in computer vision tasks. Our methodology involves storing the whole dataset directly in INR format on a GPU, mitigating the significant data communication overhead between the CPU and GPU during training. Additionally, the decoding process from INR to RGB format is highly parallelized and executed on-the-fly. To further enhance compression, we propose iterative and dynamic pruning, as well as layer-wise quantization, building upon previous work. We evaluate our framework on the image classification task, utilizing the ResNet-18 backbone network and three commonly used datasets with varying image sizes. Rapid-INR reduces memory consumption to only about 5% of the original dataset size in RGB format and achieves a maximum 6$\times$ speedup over the PyTorch training pipeline, as well as a maximum 1.2x speedup over the DALI training pipeline, with only a marginal decrease in accuracy. Importantly, Rapid-INR can be readily applied to other computer vision tasks and backbone networks with reasonable engineering efforts. Our implementation code is publicly available at https://github.com/sharc-lab/Rapid-INR.
Template matching is a fundamental problem in computer vision with applications in fields including object detection, image registration, and object tracking. Current methods rely on nearest-neighbour (NN) matching, where the query feature space is converted to NN space by representing each query pixel with its NN in the template. NN-based methods have been shown to perform better in occlusions, appearance changes, and non-rigid transformations; however, they scale poorly with high-resolution data and high feature dimensions. We present an NN-based method which efficiently reduces the NN computations and introduces filtering in the NN fields (NNFs). A vector quantization step is introduced before the NN calculation to represent the template with $k$ features, and the filter response over the NNFs is used to compare the template and query distributions over the features. We show that state-of-the-art performance is achieved in low-resolution data, and our method outperforms previous methods at higher resolution.
Multiple Description Coding (MDC) is an error-resilient source coding method designed for transmission over noisy channels. We present a novel MDC scheme employing a neural network based on implicit neural representation. This involves overfitting the neural representation for images. Each description is transmitted along with model parameters and its respective latent spaces. Our method has advantages over traditional MDC that utilizes auto-encoders, such as eliminating the need for model training and offering high flexibility in redundancy adjustment. Experiments demonstrate that our solution is competitive with autoencoder-based MDC and classic MDC based on HEVC, delivering superior visual quality.
We present QNNRepair, the first method in the literature for repairing quantized neural networks (QNNs). QNNRepair aims to improve the accuracy of a neural network model after quantization. It accepts the full-precision and weight-quantized neural networks and a repair dataset of passing and failing tests. At first, QNNRepair applies a software fault localization method to identify the neurons that cause performance degradation during neural network quantization. Then, it formulates the repair problem into a linear programming problem of solving neuron weights parameters, which corrects the QNN's performance on failing tests while not compromising its performance on passing tests. We evaluate QNNRepair with widely used neural network architectures such as MobileNetV2, ResNet, and VGGNet on popular datasets, including high-resolution images. We also compare QNNRepair with the state-of-the-art data-free quantization method SQuant. According to the experiment results, we conclude that QNNRepair is effective in improving the quantized model's performance in most cases. Its repaired models have 24% higher accuracy than SQuant's in the independent validation set, especially for the ImageNet dataset.
Compactness in deep learning can be critical to a model's viability in low-resource applications, and a common approach to extreme model compression is quantization. We consider Iterative Product Quantization (iPQ) with Quant-Noise to be state-of-the-art in this area, but this quantization framework suffers from preventable inference quality degradation due to prevalent empty clusters. In this paper, we propose several novel enhancements aiming to improve the accuracy of iPQ with Quant-Noise by focusing on resolving empty clusters. Our contribution, which we call Partitioning-Guided k-means (PG k-means), is a heavily augmented k-means implementation composed of three main components. First, we propose a partitioning-based pre-assignment strategy that ensures no initial empty clusters and encourages an even weight-to-cluster distribution. Second, we propose an empirically superior empty cluster resolution heuristic executed via cautious partitioning of large clusters. Finally, we construct an optional optimization step that consolidates intuitively dense clusters of weights to ensure shared representation. The proposed approach consistently reduces the number of empty clusters in iPQ with Quant-Noise by 100x on average, uses 8x fewer iterations during empty cluster resolution, and improves overall model accuracy by up to 12%, when applied to RoBERTa on a variety of tasks in the GLUE benchmark.