Embedding tables are used by machine learning systems to work with categorical features. In modern Recommendation Systems, these tables can be very large, necessitating the development of new methods for fitting them in memory, even during training. We suggest Clustered Compositional Embeddings (CCE) which combines clustering-based compression like quantization to codebooks with dynamic methods like The Hashing Trick and Compositional Embeddings (Shi et al., 2020). Experimentally CCE achieves the best of both worlds: The high compression rate of codebook-based quantization, but *dynamically* like hashing-based methods, so it can be used during training. Theoretically, we prove that CCE is guaranteed to converge to the optimal codebook and give a tight bound for the number of iterations required.
Deep learning model compression is an improving and important field for the edge deployment of deep learning models. Given the increasing size of the models and their corresponding power consumption, it is vital to decrease the model size and compute requirement without a significant drop in the model's performance. In this paper, we present model compression algorithms for both non-retraining and retraining conditions. In the first case where retraining of the model is not feasible due to lack of access to the original data or absence of necessary compute resources while only having access to off-the-shelf models, we propose the Bin & Quant algorithm for compression of the deep learning models using the sensitivity of the network parameters. This results in 13x compression of the speech command and control model and 7x compression of the DeepSpeech2 models. In the second case when the models can be retrained and utmost compression is required for the negligible loss in accuracy, we propose our novel gradient-weighted k-means clustering algorithm (GWK). This method uses the gradients in identifying the important weight values in a given cluster and nudges the centroid towards those values, thereby giving importance to sensitive weights. Our method effectively combines product quantization with the EWGS[1] algorithm for sub-1-bit representation of the quantized models. We test our GWK algorithm on the CIFAR10 dataset across a range of models such as ResNet20, ResNet56, MobileNetv2 and show 35x compression on quantized models for less than 2% absolute loss in accuracy compared to the floating-point models.
Inverted file structure is a common technique for accelerating dense retrieval. It clusters documents based on their embeddings; during searching, it probes nearby clusters w.r.t. an input query and only evaluates documents within them by subsequent codecs, thus avoiding the expensive cost of exhaustive traversal. However, the clustering is always lossy, which results in the miss of relevant documents in the probed clusters and hence degrades retrieval quality. In contrast, lexical matching, such as overlaps of salient terms, tends to be strong feature for identifying relevant documents. In this work, we present the Hybrid Inverted Index (HI$^2$), where the embedding clusters and salient terms work collaboratively to accelerate dense retrieval. To make best of both effectiveness and efficiency, we devise a cluster selector and a term selector, to construct compact inverted lists and efficiently searching through them. Moreover, we leverage simple unsupervised algorithms as well as end-to-end knowledge distillation to learn these two modules, with the latter further boosting the effectiveness. Based on comprehensive experiments on popular retrieval benchmarks, we verify that clusters and terms indeed complement each other, enabling HI$^2$ to achieve lossless retrieval quality with competitive efficiency across various index settings. Our code and checkpoint are publicly available at https://github.com/namespace-Pt/Adon/tree/HI2.
There are growing interests in adapting large-scale language models using parameter-efficient fine-tuning methods. However, accelerating the model itself and achieving better inference efficiency through model compression has not been thoroughly explored yet. Model compression could provide the benefits of reducing memory footprints, enabling low-precision computations, and ultimately achieving cost-effective inference. To combine parameter-efficient adaptation and model compression, we propose AlphaTuning consisting of post-training quantization of the pre-trained language model and fine-tuning only some parts of quantized parameters for a target task. Specifically, AlphaTuning works by employing binary-coding quantization, which factorizes the full-precision parameters into binary parameters and a separate set of scaling factors. During the adaptation phase, the binary values are frozen for all tasks, while the scaling factors are fine-tuned for the downstream task. We demonstrate that AlphaTuning, when applied to GPT-2 and OPT, performs competitively with full fine-tuning on a variety of downstream tasks while achieving >10x compression ratio under 4-bit quantization and >1,000x reduction in the number of trainable parameters.
Rendering high-resolution (HR) graphics brings substantial computational costs. Efficient graphics super-resolution (SR) methods may achieve HR rendering with small computing resources and have attracted extensive research interests in industry and research communities. We present a new method for real-time SR for computer graphics, namely Super-Resolution by Predicting Offsets (SRPO). Our algorithm divides the image into two parts for processing, i.e., sharp edges and flatter areas. For edges, different from the previous SR methods that take the anti-aliased images as inputs, our proposed SRPO takes advantage of the characteristics of rasterized images to conduct SR on the rasterized images. To complement the residual between HR and low-resolution (LR) rasterized images, we train an ultra-efficient network to predict the offset maps to move the appropriate surrounding pixels to the new positions. For flat areas, we found simple interpolation methods can already generate reasonable output. We finally use a guided fusion operation to integrate the sharp edges generated by the network and flat areas by the interpolation method to get the final SR image. The proposed network only contains 8,434 parameters and can be accelerated by network quantization. Extensive experiments show that the proposed SRPO can achieve superior visual effects at a smaller computational cost than the existing state-of-the-art methods.
We present a deep learning approach for learning the joint semantic embeddings of images and captions in a Euclidean space, such that the semantic similarity is approximated by the L2 distances in the embedding space. For that, we introduce a metric learning scheme that utilizes multitask learning to learn the embedding of identical semantic concepts using a center loss. By introducing a differentiable quantization scheme into the end-to-end trainable network, we derive a semantic embedding of semantically similar concepts in Euclidean space. We also propose a novel metric learning formulation using an adaptive margin hinge loss, that is refined during the training phase. The proposed scheme was applied to the MS-COCO, Flicke30K and Flickr8K datasets, and was shown to compare favorably with contemporary state-of-the-art approaches.
Quantizing a Deep Neural Network (DNN) model to be used on a custom accelerator with efficient fixed-point hardware implementations, requires satisfying many stringent hardware-friendly quantization constraints to train the model. We evaluate the two main classes of hardware-friendly quantization methods in the context of weight quantization: the traditional Mean Squared Quantization Error (MSQE)-based methods and the more recent gradient-based methods. We study the two methods on MobileNetV1 and MobileNetV2 using multiple empirical metrics to identify the sources of performance differences between the two classes, namely, sensitivity to outliers and convergence instability of the quantizer scaling factor. Using those insights, we propose various techniques to improve the performance of both quantization methods - they fix the optimization instability issues present in the MSQE-based methods during quantization of MobileNet models and allow us to improve validation performance of the gradient-based methods by 4.0% and 3.3% for MobileNetV1 and MobileNetV2 on ImageNet respectively.
Sets have been used for modeling various types of objects (e.g., a document as the set of keywords in it and a customer as the set of the items that she has purchased). Measuring similarity (e.g., Jaccard Index) between sets has been a key building block of a wide range of applications, including, plagiarism detection, recommendation, and graph compression. However, as sets have grown in numbers and sizes, the computational cost and storage required for set similarity computation have become substantial, and this has led to the development of hashing and sketching based solutions. In this work, we propose Set2Box, a learning-based approach for compressed representations of sets from which various similarity measures can be estimated accurately in constant time. The key idea is to represent sets as boxes to precisely capture overlaps of sets. Additionally, based on the proposed box quantization scheme, we design Set2Box+, which yields more concise but more accurate box representations of sets. Through extensive experiments on 8 real-world datasets, we show that, compared to baseline approaches, Set2Box+ is (a) Accurate: achieving up to 40.8X smaller estimation error while requiring 60% fewer bits to encode sets, (b) Concise: yielding up to 96.8X more concise representations with similar estimation error, and (c) Versatile: enabling the estimation of four set-similarity measures from a single representation of each set.
Multimedia Forensics experts work consists in providing answers about integrity of a specific media content and from where it comes from. Exploitation of any traces from JPEG double compressed images is often one of the main investigative path to be used for these purposes. Thus it is fundamental to have tools and algorithms able to safely estimate the first quantization matrix to further proceed with camera model identification and related tasks. In this paper, a technique based on extensive simulation is proposed, with the aim to infer the first quantization for a certain numbers of Discrete Cosine Transform (DCT) coefficients exploiting local image statistics without using any a-priori knowledge. The method provides also a reliable confidence value for the estimation which is of great importance for forensic purposes. Experimental results w.r.t. the state-of-the-art demonstrate the effectiveness of the proposed technique both in terms of precision and overall reliability.
This paper addresses the consensus problem of first-order continuous-time multi-agent systems over undirected graphs. Each agent samples relative state measurements in a self-triggered fashion and transmits the sum of the measurements to its neighbors. Moreover, we use finite-level dynamic quantizers and apply the zooming-in technique. The proposed joint design method for quantization and self-triggered sampling achieves asymptotic consensus, and inter-event times are strictly positive. Sampling times are determined explicitly with iterative procedures including the computation of the Lambert $W$-function. A simulation example is provided to illustrate the effectiveness of the proposed method.
Weight and activation binarization can efficiently compress deep neural networks and accelerate model inference, but cause severe accuracy degradation. Existing optimization methods for binary neural networks (BNNs) focus on fitting full-precision networks to reduce quantization errors, and suffer from the trade-off between accuracy and computational complexity. In contrast, considering the limited learning ability and information loss caused by the limited representational capability of BNNs, we propose IR$^2$Net to stimulate the potential of BNNs and improve the network accuracy by restricting the input information and recovering the feature information, including: 1) information restriction: for a BNN, by evaluating the learning ability on the input information, discarding some of the information it cannot focus on, and limiting the amount of input information to match its learning ability; 2) information recovery: due to the information loss in forward propagation, the output feature information of the network is not enough to support accurate classification. By selecting some shallow feature maps with richer information, and fusing them with the final feature maps to recover the feature information. In addition, the computational cost is reduced by streamlining the information recovery method to strike a better trade-off between accuracy and efficiency. Experimental results demonstrate that our approach still achieves comparable accuracy even with $ \sim $10x floating-point operations (FLOPs) reduction for ResNet-18. The models and code are available at https://github.com/pingxue-hfut/IR2Net.
We introduce GLM-130B, a bilingual (English and Chinese) pre-trained language model with 130 billion parameters. It is an attempt to open-source a 100B-scale model at least as good as GPT-3 (davinci) and unveil how models of such a scale can be successfully pre-trained. Over the course of this effort, we face numerous unexpected technical and engineering challenges, particularly on loss spikes and divergence. In this paper, we introduce the training process of GLM-130B including its design choices, training strategies for both efficiency and stability, and engineering efforts. The resultant GLM-130B model offers significant outperformance over GPT-3 175B (davinci) on a wide range of popular English benchmarks while the performance advantage is not observed in OPT-175B and BLOOM-176B. It also consistently and significantly outperforms ERNIE TITAN 3.0 260B -- the largest Chinese language model -- across related benchmarks. Finally, we leverage a unique scaling property of GLM-130B to reach INT4 quantization without post training, with almost no performance loss, making it the first among 100B-scale models and more importantly, allowing its effective inference on 4$\times$RTX 3090 (24G) or 8$\times$RTX 2080 Ti (11G) GPUs, the most affordable GPUs required for using 100B-scale models. The GLM-130B model weights are publicly accessible and its code, training logs, related toolkit, and lessons learned are open-sourced at \url{https://github.com/THUDM/GLM-130B/}.
Accumulation of corporate data in the cloud has attracted more enterprise applications to the cloud creating data gravity. As a consequence, network traffic has become more cloud centric. This increase in cloud centric traffic poses new challenges in designing learning systems for streaming data due to class imbalance. The number of classes plays a vital role in the accuracy of the classifiers built from the data streams. In this paper, we present a vector quantization-based sampling method, which substantially reduces the class imbalance in data streams. We demonstrate its effectiveness by conducting experiments on network traffic and anomaly dataset with commonly used ML model building methods; Multilayered Perceptron on TensorFlow backend, Support Vector Machines, K-Nearest Neighbour, and Random Forests. We built models using parallel processing, batch processing, and randomly selecting samples. We show that the accuracy of classification models improves when the data streams are pre-processed with our method. We used out of the box hyper-parameters of these classifiers and auto sklearn for hyperparameter optimization.
Light-field displays create an immersive experience by providing binocular depth sensation and motion parallax. Stacking light attenuating layers is one approach to implement a light field display with a broader depth of field, wide viewing angles and high resolution. Due to the transparent holographic optical element (HOE) layers, additive layered displays can be integrated into augmented reality (AR) wearables to overlay virtual objects onto the real world, creating a seamless mixed reality (XR) experience. This paper proposes a novel framework for light field representation and coding that utilizes Deep Belief Network (DBN) and weighted binary images suitable for additive layered displays. The weighted binary representation of layers makes the framework more flexible for adaptive bitrate encoding. The framework effectively captures intrinsic redundancies in the light field data, and thus provides a scalable solution for light field coding suitable for XR display applications. The latent code is encoded by H.265 codec generating a rate-scalable bit-stream. We achieve adaptive bitrate decoding by varying the number of weighted binary images and the H.265 quantization parameter, while maintaining an optimal reconstruction quality. The framework is tested on real and synthetic benchmark datasets, and the results validate the rate-scalable property of the proposed scheme.
In this paper, the situation in which a receiver has to execute a task from a quantized version of the information source of interest is considered. The task is modeled by the minimization problem of a general goal function $f(x;g)$ for which the decision $x$ has to be taken from a quantized version of the parameters $g$. This problem is relevant in many applications e.g., for radio resource allocation (RA), high spectral efficiency communications, controlled systems, or data clustering in the smart grid. By resorting to high resolution (HR) analysis, it is shown how to design a quantizer that minimizes the gap between the minimum of $f$ (which would be reached by knowing $g$ perfectly) and what is effectively reached with a quantized $g$. The conducted formal analysis both provides quantization strategies in the HR regime and insights for the general regime and allows a practical algorithm to be designed. The analysis also allows one to provide some elements to the new and fundamental problem of the relationship between the goal function regularity properties and the hardness to quantize its parameters. The derived results are discussed and supported by a rich numerical performance analysis in which known RA goal functions are studied and allows one to exhibit very significant improvements by tailoring the quantization operation to the final task.
It has been proven that, compared to using 32-bit floating-point numbers in the training phase, Deep Convolutional Neural Networks (DCNNs) can operate with low precision during inference, thereby saving memory space and power consumption. However, quantizing networks is always accompanied by an accuracy decrease. Here, we propose a method, double-stage Squeeze-and-Threshold (double-stage ST). It uses the attention mechanism to quantize networks and achieve state-of-art results. Using our method, the 3-bit model can achieve accuracy that exceeds the accuracy of the full-precision baseline model. The proposed double-stage ST activation quantization is easy to apply: inserting it before the convolution.
Smart meter measurements, though critical for accurate demand forecasting, face several drawbacks including consumers' privacy, data breach issues, to name a few. Recent literature has explored Federated Learning (FL) as a promising privacy-preserving machine learning alternative which enables collaborative learning of a model without exposing private raw data for short term load forecasting. Despite its virtue, standard FL is still vulnerable to an intractable cyber threat known as Byzantine attack carried out by faulty and/or malicious clients. Therefore, to improve the robustness of federated short-term load forecasting against Byzantine threats, we develop a state-of-the-art differentially private secured FL-based framework that ensures the privacy of the individual smart meter's data while protect the security of FL models and architecture. Our proposed framework leverages the idea of gradient quantization through the Sign Stochastic Gradient Descent (SignSGD) algorithm, where the clients only transmit the `sign' of the gradient to the control centre after local model training. As we highlight through our experiments involving benchmark neural networks with a set of Byzantine attack models, our proposed approach mitigates such threats quite effectively and thus outperforms conventional Fed-SGD models.
With the development of deep learning, neural network-based speech enhancement (SE) models have shown excellent performance. Meanwhile, it was shown that the development of self-supervised pre-trained models can be applied to various downstream tasks. In this paper, we will consider the application of the pre-trained model to the real-time SE problem. Specifically, the encoder and bottleneck layer of the DEMUCS model are initialized using the self-supervised pretrained WavLM model, the convolution in the encoder is replaced by causal convolution, and the transformer encoder in the bottleneck layer is based on causal attention mask. In addition, as discretizing the noisy speech representations is more beneficial for denoising, we utilize a quantization module to discretize the representation output from the bottleneck layer, which is then fed into the decoder to reconstruct the clean speech waveform. Experimental results on the Valentini dataset and an internal dataset show that the pre-trained model based initialization can improve the SE performance and the discretization operation suppresses the noise component in the representations to some extent, which can further improve the performance.
This paper considers the problem of lossy neural image compression (NIC). Current state-of-the-art (sota) methods adopt uniform posterior to approximate quantization noise, and single-sample pathwise estimator to approximate the gradient of evidence lower bound (ELBO). In this paper, we propose to train NIC with multiple-sample importance weighted autoencoder (IWAE) target, which is tighter than ELBO and converges to log likelihood as sample size increases. First, we identify that the uniform posterior of NIC has special properties, which affect the variance and bias of pathwise and score function estimators of the IWAE target. Moreover, we provide insights on a commonly adopted trick in NIC from gradient variance perspective. Based on those analysis, we further propose multiple-sample NIC (MS-NIC), an enhanced IWAE target for NIC. Experimental results demonstrate that it improves sota NIC methods. Our MS-NIC is plug-and-play, and can be easily extended to other neural compression tasks.
Diffractive optical neural networks (DONNs) have attracted lots of attention as they bring significant advantages in terms of power efficiency, parallelism, and computational speed compared with conventional deep neural networks (DNNs), which have intrinsic limitations when implemented on digital platforms. However, inversely mapping algorithm-trained physical model parameters onto real-world optical devices with discrete values is a non-trivial task as existing optical devices have non-unified discrete levels and non-monotonic properties. This work proposes a novel device-to-system hardware-software codesign framework, which enables efficient physics-aware training of DONNs w.r.t arbitrary experimental measured optical devices across layers. Specifically, Gumbel-Softmax is employed to enable differentiable discrete mapping from real-world device parameters into the forward function of DONNs, where the physical parameters in DONNs can be trained by simply minimizing the loss function of the ML task. The results have demonstrated that our proposed framework offers significant advantages over conventional quantization-based methods, especially with low-precision optical devices. Finally, the proposed algorithm is fully verified with physical experimental optical systems in low-precision settings.