After being trained, classifiers must often operate on data that has been corrupted by noise. In this paper, we consider the impact of such noise on the features of binary classifiers. Inspired by tools for classifier robustness, we introduce the same classification probability (SCP) to measure the resulting distortion on the classifier outputs. We introduce a low-complexity estimate of the SCP based on quantization and polynomial multiplication. We also study channel coding techniques based on replication error-correcting codes. In contrast to the traditional channel coding approach, where error-correction is meant to preserve the data and is agnostic to the application, our schemes specifically aim to maximize the SCP (equivalently minimizing the distortion of the classifier output) for the same redundancy overhead.
Guetzli is a new JPEG encoder that aims to produce visually indistinguishable images at a lower bit-rate than other common JPEG encoders. It optimizes both the JPEG global quantization tables and the DCT coefficient values in each JPEG block using a closed-loop optimizer. Guetzli uses Butteraugli, our perceptual distance metric, as the source of feedback in its optimization process. We reach a 29-45% reduction in data size for a given perceptual distance, according to Butteraugli, in comparison to other compressors we tried. Guetzli's computation is currently extremely slow, which limits its applicability to compressing static content and serving as a proof- of-concept that we can achieve significant reductions in size by combining advanced psychovisual models with lossy compression techniques.
The traditional bag-of-words approach has found a wide range of applications in computer vision. The standard pipeline consists of a generation of a visual vocabulary, a quantization of the features into histograms of visual words, and a classification step for which usually a support vector machine in combination with a non-linear kernel is used. Given large amounts of data, however, the model suffers from a lack of discriminative power. This applies particularly for action recognition, where the vast amount of video features needs to be subsampled for unsupervised visual vocabulary generation. Moreover, the kernel computation can be very expensive on large datasets. In this work, we propose a recurrent neural network that is equivalent to the traditional bag-of-words approach but enables for the application of discriminative training. The model further allows to incorporate the kernel computation into the neural network directly, solving the complexity issue and allowing to represent the complete classification system within a single network. We evaluate our method on four recent action recognition benchmarks and show that the conventional model as well as sparse coding methods are outperformed.
The nearest neighbor method together with the dynamic time warping (DTW) distance is one of the most popular approaches in time series classification. This method suffers from high storage and computation requirements for large training sets. As a solution to both drawbacks, this article extends learning vector quantization (LVQ) from Euclidean spaces to DTW spaces. The proposed LVQ scheme uses asymmetric weighted averaging as update rule. Empirical results exhibited superior performance of asymmetric generalized LVQ (GLVQ) over other state-of-the-art prototype generation methods for nearest neighbor classification.
This paper explores the use of Pyramid Vector Quantization (PVQ) to reduce the computational cost for a variety of neural networks (NNs) while, at the same time, compressing the weights that describe them. This is based on the fact that the dot product between an N dimensional vector of real numbers and an N dimensional PVQ vector can be calculated with only additions and subtractions and one multiplication. This is advantageous since tensor products, commonly used in NNs, can be re-conduced to a dot product or a set of dot products. Finally, it is stressed that any NN architecture that is based on an operation that can be re-conduced to a dot product can benefit from the techniques described here.
For computer vision applications, prior works have shown the efficacy of reducing the numeric precision of model parameters (network weights) in deep neural networks but also that reducing the precision of activations hurts model accuracy much more than reducing the precision of model parameters. We study schemes to train networks from scratch using reduced-precision activations without hurting the model accuracy. We reduce the precision of activation maps (along with model parameters) using a novel quantization scheme and increase the number of filter maps in a layer, and find that this scheme compensates or surpasses the accuracy of the baseline full-precision network. As a result, one can significantly reduce the dynamic memory footprint, memory bandwidth, computational energy and speed up the training and inference process with appropriate hardware support. We call our scheme WRPN - wide reduced-precision networks. We report results using our proposed schemes and show that our results are better than previously reported accuracies on ILSVRC-12 dataset while being computationally less expensive compared to previously reported reduced-precision networks.
We formulate the notion of minimax estimation under storage or communication constraints, and prove an extension to Pinsker's theorem for nonparametric estimation over Sobolev ellipsoids. Placing limits on the number of bits used to encode any estimator, we give tight lower and upper bounds on the excess risk due to quantization in terms of the number of bits, the signal size, and the noise level. This establishes the Pareto optimal tradeoff between storage and risk under quantization constraints for Sobolev spaces. Our results and proof techniques combine elements of rate distortion theory and minimax analysis. The proposed quantized estimation scheme, which shows achievability of the lower bounds, is adaptive in the usual statistical sense, achieving the optimal quantized minimax rate without knowledge of the smoothness parameter of the Sobolev space. It is also adaptive in a computational sense, as it constructs the code only after observing the data, to dynamically allocate more codewords to blocks where the estimated signal size is large. Simulations are included that illustrate the effect of quantization on statistical risk.
In this paper, we propose a product quantization table (PQTable); a fast search method for product-quantized codes via hash-tables. An identifier of each database vector is associated with the slot of a hash table by using its PQ-code as a key. For querying, an input vector is PQ-encoded and hashed, and the items associated with that code are then retrieved. The proposed PQTable produces the same results as a linear PQ scan, and is 10^2 to 10^5 times faster. Although state-of-the-art performance can be achieved by previous inverted-indexing-based approaches, such methods require manually-designed parameter setting and significant training; our PQTable is free of these limitations, and therefore offers a practical and effective solution for real-world problems. Specifically, when the vectors are highly compressed, our PQTable achieves one of the fastest search performances on a single CPU to date with significantly efficient memory usage (0.059 ms per query over 10^9 data points with just 5.5 GB memory consumption). Finally, we show that our proposed PQTable can naturally handle the codes of an optimized product quantization (OPQTable).
Nearest neighbor search is a problem of finding the data points from the database such that the distances from them to the query point are the smallest. Learning to hash is one of the major solutions to this problem and has been widely studied recently. In this paper, we present a comprehensive survey of the learning to hash algorithms, categorize them according to the manners of preserving the similarities into: pairwise similarity preserving, multiwise similarity preserving, implicit similarity preserving, as well as quantization, and discuss their relations. We separate quantization from pairwise similarity preserving as the objective function is very different though quantization, as we show, can be derived from preserving the pairwise similarities. In addition, we present the evaluation protocols, and the general performance analysis, and point out that the quantization algorithms perform superiorly in terms of search accuracy, search time cost, and space cost. Finally, we introduce a few emerging topics.
Efficient Nearest Neighbor (NN) search in high-dimensional spaces is a foundation of many multimedia retrieval systems. Because it offers low responses times, Product Quantization (PQ) is a popular solution. PQ compresses high-dimensional vectors into short codes using several sub-quantizers, which enables in-RAM storage of large databases. This allows fast answers to NN queries, without accessing the SSD or HDD. The key feature of PQ is that it can compute distances between short codes and high-dimensional vectors using cache-resident lookup tables. The efficiency of this technique, named Asymmetric Distance Computation (ADC), remains limited because it performs many cache accesses. In this paper, we introduce Quick ADC, a novel technique that achieves a 3 to 6 times speedup over ADC by exploiting Single Instruction Multiple Data (SIMD) units available in current CPUs. Efficiently exploiting SIMD requires algorithmic changes to the ADC procedure. Namely, Quick ADC relies on two key modifications of ADC: (i) the use 4-bit sub-quantizers instead of the standard 8-bit sub-quantizers and (ii) the quantization of floating-point distances. This allows Quick ADC to exceed the performance of state-of-the-art systems, e.g., it achieves a Recall@100 of 0.94 in 3.4 ms on 1 billion SIFT descriptors (128-bit codes).
The Residual Quantization (RQ) framework is revisited where the quantization distortion is being successively reduced in multi-layers. Inspired by the reverse-water-filling paradigm in rate-distortion theory, an efficient regularization on the variances of the codewords is introduced which allows to extend the RQ for very large numbers of layers and also for high dimensional data, without getting over-trained. The proposed Regularized Residual Quantization (RRQ) results in multi-layer dictionaries which are additionally sparse, thanks to the soft-thresholding nature of the regularization when applied to variance-decaying data which can arise from de-correlating transformations applied to correlated data. Furthermore, we also propose a general-purpose pre-processing for natural images which makes them suitable for such quantization. The RRQ framework is first tested on synthetic variance-decaying data to show its efficiency in quantization of high-dimensional data. Next, we use the RRQ in super-resolution of a database of facial images where it is shown that low-resolution facial images from the test set quantized with codebooks trained on high-resolution images from the training set show relevant high-frequency content when reconstructed with those codebooks.
While deep learning has become a key ingredient in the top performing methods for many computer vision tasks, it has failed so far to bring similar improvements to instance-level image retrieval. In this article, we argue that reasons for the underwhelming results of deep methods on image retrieval are threefold: i) noisy training data, ii) inappropriate deep architecture, and iii) suboptimal training procedure. We address all three issues. First, we leverage a large-scale but noisy landmark dataset and develop an automatic cleaning method that produces a suitable training set for deep retrieval. Second, we build on the recent R-MAC descriptor, show that it can be interpreted as a deep and differentiable architecture, and present improvements to enhance it. Last, we train this network with a siamese architecture that combines three streams with a triplet loss. At the end of the training process, the proposed architecture produces a global image representation in a single forward pass that is well suited for image retrieval. Extensive experiments show that our approach significantly outperforms previous retrieval approaches, including state-of-the-art methods based on costly local descriptor indexing and spatial verification. On Oxford 5k, Paris 6k and Holidays, we respectively report 94.7, 96.6, and 94.8 mean average precision. Our representations can also be heavily compressed using product quantization with little loss in accuracy. For additional material, please see www.xrce.xerox.com/Deep-Image-Retrieval.
It is of fundamental importance to find algorithms obtaining optimal performance for learning of statistical models in distributed and communication limited systems. Aiming at characterizing the optimal strategies, we consider learning of Gaussian Processes (GPs) in distributed systems as a pivotal example. We first address a very basic problem: how many bits are required to estimate the inner-products of Gaussian vectors across distributed machines? Using information theoretic bounds, we obtain an optimal solution for the problem which is based on vector quantization. Two suboptimal and more practical schemes are also presented as substitute for the vector quantization scheme. In particular, it is shown that the performance of one of the practical schemes which is called per-symbol quantization is very close to the optimal one. Schemes provided for the inner-product calculations are incorporated into our proposed distributed learning methods for GPs. Experimental results show that with spending few bits per symbol in our communication scheme, our proposed methods outperform previous zero rate distributed GP learning schemes such as Bayesian Committee Model (BCM) and Product of experts (PoE).
The success of deep learning in numerous application domains created the de- sire to run and train them on mobile devices. This however, conflicts with their computationally, memory and energy intense nature, leading to a growing interest in compression. Recent work by Han et al. (2015a) propose a pipeline that involves retraining, pruning and quantization of neural network weights, obtaining state-of-the-art compression rates. In this paper, we show that competitive compression rates can be achieved by using a version of soft weight-sharing (Nowlan & Hinton, 1992). Our method achieves both quantization and pruning in one simple (re-)training procedure. This point of view also exposes the relation between compression and the minimum description length (MDL) principle.
In this paper, we study the prediction of a circularly symmetric zero-mean stationary Gaussian process from a window of observations consisting of finitely many samples. This is a prevalent problem in a wide range of applications in communication theory and signal processing. Due to stationarity, when the autocorrelation function or equivalently the power spectral density (PSD) of the process is available, the Minimum Mean Squared Error (MMSE) predictor is readily obtained. In particular, it is given by a linear operator that depends on autocorrelation of the process as well as the noise power in the observed samples. The prediction becomes, however, quite challenging when the PSD of the process is unknown. In this paper, we propose a blind predictor that does not require the a priori knowledge of the PSD of the process and compare its performance with that of an MMSE predictor that has a full knowledge of the PSD. To design such a blind predictor, we use the random spectral representation of a stationary Gaussian process. We apply the well-known atomic-norm minimization technique to the observed samples to obtain a discrete quantization of the underlying random spectrum, which we use to predict the process. Our simulation results show that this estimator has a good performance comparable with that of the MMSE estimator.
Deep convolutional Neural Networks (CNN) are the state-of-the-art performers for object detection task. It is well known that object detection requires more computation and memory than image classification. Thus the consolidation of a CNN-based object detection for an embedded system is more challenging. In this work, we propose LCDet, a fully-convolutional neural network for generic object detection that aims to work in embedded systems. We design and develop an end-to-end TensorFlow(TF)-based model. Additionally, we employ 8-bit quantization on the learned weights. We use face detection as a use case. Our TF-Slim based network can predict different faces of different shapes and sizes in a single forward pass. Our experimental results show that the proposed method achieves comparative accuracy comparing with state-of-the-art CNN-based face detection methods, while reducing the model size by 3x and memory-BW by ~4x comparing with one of the best real-time CNN-based object detector such as YOLO. TF 8-bit quantized model provides additional 4x memory reduction while keeping the accuracy as good as the floating point model. The proposed model thus becomes amenable for embedded implementations.
We propose the residual expansion (RE) algorithm: a global (or near-global) optimization method for nonconvex least squares problems. Unlike most existing nonconvex optimization techniques, the RE algorithm is not based on either stochastic or multi-point searches; therefore, it can achieve fast global optimization. Moreover, the RE algorithm is easy to implement and successful in high-dimensional optimization. The RE algorithm exhibits excellent empirical performance in terms of k-means clustering, point-set registration, optimized product quantization, and blind image deblurring.
Data quantization learns encoding results of data with certain requirements, and provides a broad perspective of many real-world applications to data handling. Nevertheless, the results of encoder is usually limited to multivariate inputs with the random mapping, and side information of binary codes are hardly to mostly depict the original data patterns as possible. In the literature, cosine based random quantization has attracted much attentions due to its intrinsic bounded results. Nevertheless, it usually suffers from the uncertain outputs, and information of original data fails to be fully preserved in the reduced codes. In this work, a novel binary embedding method, termed adaptive training quantization (ATQ), is proposed to learn the ideal transform of random encoder, where the limitation of cosine random mapping is tackled. As an adaptive learning idea, the reduced mapping is adaptively calculated with idea of data group, while the bias of random transform is to be improved to hold most matching information. Experimental results show that the proposed method is able to obtain outstanding performance compared with other random quantization methods.
We propose a novel fine-grained quantization (FGQ) method to ternarize pre-trained full precision models, while also constraining activations to 8 and 4-bits. Using this method, we demonstrate a minimal loss in classification accuracy on state-of-the-art topologies without additional training. We provide an improved theoretical formulation that forms the basis for a higher quality solution using FGQ. Our method involves ternarizing the original weight tensor in groups of $N$ weights. Using $N=4$, we achieve Top-1 accuracy within $3.7\%$ and $4.2\%$ of the baseline full precision result for Resnet-101 and Resnet-50 respectively, while eliminating $75\%$ of all multiplications. These results enable a full 8/4-bit inference pipeline, with best-reported accuracy using ternary weights on ImageNet dataset, with a potential of $9\times$ improvement in performance. Also, for smaller networks like AlexNet, FGQ achieves state-of-the-art results. We further study the impact of group size on both performance and accuracy. With a group size of $N=64$, we eliminate $\approx99\%$ of the multiplications; however, this introduces a noticeable drop in accuracy, which necessitates fine tuning the parameters at lower precision. We address this by fine-tuning Resnet-50 with 8-bit activations and ternary weights at $N=64$, improving the Top-1 accuracy to within $4\%$ of the full precision result with $<30\%$ additional training overhead. Our final quantized model can run on a full 8-bit compute pipeline using 2-bit weights and has the potential of up to $15\times$ improvement in performance compared to baseline full-precision models.
Multimodal biometric identification has been grown a great attention in the most interests in the security fields. In the real world there exist modern system devices that are able to detect, recognize, and classify the human identities with reliable and fast recognition rates. Unfortunately most of these systems rely on one modality, and the reliability for two or more modalities are further decreased. The variations of face images with respect to different poses are considered as one of the important challenges in face recognition systems. In this paper, we propose a multimodal biometric system that able to detect the human face images that are not only one view face image, but also multi-view face images. Each subject entered to the system adjusted their face at front of the three cameras, and then the features of the face images are extracted based on Speeded Up Robust Features (SURF) algorithm. We utilize Multi-Layer Perceptron (MLP) and combined classifiers based on both Learning Vector Quantization (LVQ), and Radial Basis Function (RBF) for classification purposes. The proposed system has been tested using SDUMLA-HMT, and CASIA datasets. Furthermore, we collected a database of multi-view face images by which we take the additive white Gaussian noise into considerations. The results indicated the reliability, robustness of the proposed system with different poses and variations including noise images.