Convolutional neural networks are modern models that are very efficient in many classification tasks. They were originally created for image processing purposes. Then some trials were performed to use them in different domains like natural language processing. The artificial intelligence systems (like humanoid robots) are very often based on embedded systems with constraints on memory, power consumption etc. Therefore convolutional neural network because of its memory capacity should be reduced to be mapped to given hardware. In this paper, results are presented of compressing the efficient convolutional neural networks for sentiment analysis. The main steps are quantization and pruning processes. The method responsible for mapping compressed network to FPGA and results of this implementation are presented. The described simulations showed that 5-bit width is enough to have no drop in accuracy from floating point version of the network. Additionally, significant memory footprint reduction was achieved (from 85% up to 93%).
Model compression has gained a lot of attention due to its ability to reduce hardware resource requirements significantly while maintaining accuracy of DNNs. Model compression is especially useful for memory-intensive recurrent neural networks because smaller memory footprint is crucial not only for reducing storage requirement but also for fast inference operations. Quantization is known to be an effective model compression method and researchers are interested in minimizing the number of bits to represent parameters. In this work, we introduce an iterative technique to apply quantization, presenting high compression ratio without any modifications to the training algorithm. In the proposed technique, weight quantization is followed by retraining the model with full precision weights. We show that iterative retraining generates new sets of weights which can be quantized with decreasing quantization loss at each iteration. We also show that quantization is efficiently able to leverage pruning, another effective model compression method. Implementation issues on combining the two methods are also addressed. Our experimental results demonstrate that an LSTM model using 1-bit quantized weights is sufficient for PTB dataset without any accuracy degradation while previous methods demand at least 2-4 bits for quantized weights.
Deep neural networks (DNNs) have become the state-of-the-art technique for machine learning tasks in various applications. However, due to their size and the computational complexity, large DNNs are not readily deployable on edge devices in real-time. To manage complexity and accelerate computation, network compression techniques based on pruning and quantization have been proposed and shown to be effective in reducing network size. However, such network compression can result in irregular matrix structures that are mismatched with modern hardware-accelerated platforms, such as graphics processing units (GPUs) designed to perform the DNN matrix multiplications in a structured (block-based) way. We propose MPDCompress, a DNN compression algorithm based on matrix permutation decomposition via random mask generation. In-training application of the masks molds the synaptic weight connection matrix to a sub-graph separation format. Aided by the random permutations, a hardware-desirable block matrix is generated, allowing for a more efficient implementation and compression of the network. To show versatility, we empirically verify MPDCompress on several network models, compression rates, and image datasets. On the LeNet 300-100 model (MNIST dataset), Deep MNIST, and CIFAR10, we achieve 10 X network compression with less than 1% accuracy loss compared to non-compressed accuracy performance. On AlexNet for the full ImageNet ILSVRC-2012 dataset, we achieve 8 X network compression with less than 1% accuracy loss, with top-5 and top-1 accuracies of 79.6% and 56.4%, respectively. Finally, we observe that the algorithm can offer inference speedups across various hardware platforms, with 4 X faster operation achieved on several mobile GPUs.
Deep Neural Networks trained as image auto-encoders have recently emerged as a promising direction for advancing the state-of-the-art in image compression. The key challenge in learning such networks is twofold: To deal with quantization, and to control the trade-off between reconstruction error (distortion) and entropy (rate) of the latent image representation. In this paper, we focus on the latter challenge and propose a new technique to navigate the rate-distortion trade-off for an image compression auto-encoder. The main idea is to directly model the entropy of the latent representation by using a context model: A 3D-CNN which learns a conditional probability model of the latent distribution of the auto-encoder. During training, the auto-encoder makes use of the context model to estimate the entropy of its representation, and the context model is concurrently updated to learn the dependencies between the symbols in the latent representation. Our experiments show that this approach, when measured in MS-SSIM, yields a state-of-the-art image compression system based on a simple convolutional auto-encoder.
Recently, many studies have demonstrated deep neural network (DNN) classifiers can be fooled by the adversarial example, which is crafted via introducing some perturbations into an original sample. Accordingly, some powerful defense techniques were proposed. However, existing defense techniques often require modifying the target model or depend on the prior knowledge of attacks. In this paper, we propose a straightforward method for detecting adversarial image examples, which can be directly deployed into unmodified off-the-shelf DNN models. We consider the perturbation to images as a kind of noise and introduce two classic image processing techniques, scalar quantization and smoothing spatial filter, to reduce its effect. The image entropy is employed as a metric to implement an adaptive noise reduction for different kinds of images. Consequently, the adversarial example can be effectively detected by comparing the classification results of a given sample and its denoised version, without referring to any prior knowledge of attacks. More than 20,000 adversarial examples against some state-of-the-art DNN models are used to evaluate the proposed method, which are crafted with different attack techniques. The experiments show that our detection method can achieve a high overall F1 score of 96.39% and certainly raises the bar for defense-aware attacks.
Most image deblurring methods assume an over-simplistic image formation model and as a result are sensitive to more realistic image degradations. We propose a novel variational framework, that explicitly handles pixel saturation, noise, quantization, as well as non-linear camera response function due to e.g., gamma correction. We show that accurately modeling a more realistic image acquisition pipeline leads to significant improvements, both in terms of image quality and PSNR. Furthermore, we show that incorporating the non-linear response in both the data and the regularization terms of the proposed energy leads to a more detailed restoration than a naive inversion of the non-linear curve. The minimization of the proposed energy is performed using stochastic optimization. A dataset consisting of realistically degraded images is created in order to evaluate the method.
We present a lossy image compression method based on deep convolutional neural networks (CNNs), which outperforms the existing BPG, WebP, JPEG2000 and JPEG as measured via multi-scale structural similarity (MS-SSIM), at the same bit rate. Currently, most of the CNNs based approaches train the network using a L2 loss between the reconstructions and the ground-truths in the pixel domain, which leads to over-smoothing results and visual quality degradation especially at a very low bit rate. Therefore, we improve the subjective quality with the combination of a perception loss and an adversarial loss additionally. To achieve better rate-distortion optimization (RDO), we also introduce an easy-to-hard transfer learning when adding quantization error and rate constraint. Finally, we evaluate our method on public Kodak and the Test Dataset P/M released by the Computer Vision Lab of ETH Zurich, resulting in averaged 7.81% and 19.1% BD-rate reduction over BPG, respectively.
Deep reinforcement learning, applied to vision-based problems like Atari games, maps pixels directly to actions; internally, the deep neural network bears the responsibility of both extracting useful information and making decisions based on it. By separating the image processing from decision-making, one could better understand the complexity of each task, as well as potentially find smaller policy representations that are easier for humans to understand and may generalize better. To this end, we propose a new method for learning policies and compact state representations separately but simultaneously for policy approximation in reinforcement learning. State representations are generated by an encoder based on two novel algorithms: Increasing Dictionary Vector Quantization makes the encoder capable of growing its dictionary size over time, to address new observations as they appear in an open-ended online-learning context; Direct Residuals Sparse Coding encodes observations by disregarding reconstruction error minimization, and aiming instead for highest information inclusion. The encoder autonomously selects observations online to train on, in order to maximize code sparsity. As the dictionary size increases, the encoder produces increasingly larger inputs for the neural network: this is addressed by a variation of the Exponential Natural Evolution Strategies algorithm which adapts its probability distribution dimensionality along the run. We test our system on a selection of Atari games using tiny neural networks of only 6 to 18 neurons (depending on the game's controls). These are still capable of achieving results comparable---and occasionally superior---to state-of-the-art techniques which use two orders of magnitude more neurons.
Modern scientific instruments produce vast amounts of data, which can overwhelm the processing ability of computer systems. Lossy compression of data is an intriguing solution, but comes with its own drawbacks, such as potential signal loss, and the need for careful optimization of the compression ratio. In this work, we focus on a setting where this problem is especially acute: compressive sensing frameworks for interferometry and medical imaging. We ask the following question: can the precision of the data representation be lowered for all inputs, with recovery guarantees and practical performance? Our first contribution is a theoretical analysis of the normalized Iterative Hard Thresholding (IHT) algorithm when all input data, meaning both the measurement matrix and the observation vector are quantized aggressively. We present a variant of low precision normalized {IHT} that, under mild conditions, can still provide recovery guarantees. The second contribution is the application of our quantization framework to radio astronomy and magnetic resonance imaging. We show that lowering the precision of the data can significantly accelerate image recovery. We evaluate our approach on telescope data and samples of brain images using CPU and FPGA implementations achieving up to a 9x speed-up with negligible loss of recovery quality.
As it requires a huge number of parameters when exposed to high dimensional inputs in video detection and classification, there is a grand challenge to develop a compact yet accurate video comprehension at terminal devices. Current works focus on optimizations of video detection and classification in a separated fashion. In this paper, we introduce a video comprehension (object detection and action recognition) system for terminal devices, namely DEEPEYE. Based on You Only Look Once (YOLO), we have developed an 8-bit quantization method when training YOLO; and also developed a tensorized-compression method of Recurrent Neural Network (RNN) composed of features extracted from YOLO. The developed quantization and tensorization can significantly compress the original network model yet with maintained accuracy. Using the challenging video datasets: MOMENTS and UCF11 as benchmarks, the results show that the proposed DEEPEYE achieves 3.994x model compression rate with only 0.47% mAP decreased; and 15,047x parameter reduction and 2.87x speed-up with 16.58% accuracy improvement.
In this paper, we propose a simple and general framework for training very tiny CNNs for object detection. Due to limited representation ability, it is challenging to train very tiny networks for complicated tasks like detection. To the best of our knowledge, our method, called Quantization Mimic, is the first one focusing on very tiny networks. We utilize two types of acceleration methods: mimic and quantization. Mimic improves the performance of a student network by transfering knowledge from a teacher network. Quantization converts a full-precision network to a quantized one without large degradation of performance. If the teacher network is quantized, the search scope of the student network will be smaller. Using this feature of the quantization, we propose Quantization Mimic. It first quantizes the large network, then mimic a quantized small network. The quantization operation can help student network to better match the feature maps from teacher network. To evaluate our approach, we carry out experiments on various popular CNNs including VGG and Resnet, as well as different detection frameworks including Faster R-CNN and R-FCN. Experiments on Pascal VOC and WIDER FACE verify that our Quantization Mimic algorithm can be applied on various settings and outperforms state-of-the-art model acceleration methods given limited computing resouces.
Discretizing multi-dimensional data distributions is a fundamental step of modern indexing methods. State-of-the-art techniques learn parameters of quantizers on training data for optimal performance, thus adapting quantizers to the data. In this work, we propose to reverse this paradigm and adapt the data to the quantizer: we train a neural net which last layer forms a fixed parameter-free quantizer, such as pre-defined points of a hyper-sphere. As a proxy objective, we design and train a neural network that favors uniformity in the spherical latent space, while preserving the neighborhood structure after the mapping. We propose a new regularizer derived from the Kozachenko--Leonenko differential entropy estimator to enforce uniformity and combine it with a locality-aware triplet loss. Experiments show that our end-to-end approach outperforms most learned quantization methods, and is competitive with the state of the art on widely adopted benchmarks. Furthermore, we show that training without the quantization step results in almost no difference in accuracy, but yields a generic catalyzer that can be applied with any subsequent quantizer.
Point cloud, an efficient 3D object representation, has become popular with the development of depth sensing and 3D laser scanning techniques. It has attracted attention in various applications such as 3D tele-presence, navigation for unmanned vehicles and heritage reconstruction. The understanding of point clouds, such as point cloud segmentation, is crucial in exploiting the informative value of point clouds for such applications. Due to the irregularity of the data format, previous deep learning works often convert point clouds to regular 3D voxel grids or collections of images before feeding them into neural networks, which leads to voluminous data and quantization artifacts. In this paper, we instead propose a regularized graph convolutional neural network (RGCNN) that directly consumes point clouds. Leveraging on spectral graph theory, we treat features of points in a point cloud as signals on graph, and define the convolution over graph by Chebyshev polynomial approximation. In particular, we update the graph Laplacian matrix that describes the connectivity of features in each layer according to the corresponding learned features, which adaptively captures the structure of dynamic graphs. Further, we deploy a graph-signal smoothness prior in the loss function, thus regularizing the learning process. Experimental results on the ShapeNet part dataset show that the proposed approach significantly reduces the computational complexity while achieving competitive performance with the state of the art. Also, experiments show RGCNN is much more robust to both noise and point cloud density in comparison with other methods. We further apply RGCNN to point cloud classification and achieve competitive results on ModelNet40 dataset.
Deep neural networks have achieved impressive results in computer vision and machine learning. Unfortunately, state-of-the-art networks are extremely compute and memory intensive which makes them unsuitable for mW-devices such as IoT end-nodes. Aggressive quantization of these networks dramatically reduces the computation and memory footprint. Binary-weight neural networks (BWNs) follow this trend, pushing weight quantization to the limit. Hardware accelerators for BWNs presented up to now have focused on core efficiency, disregarding I/O bandwidth and system-level efficiency that are crucial for deployment of accelerators in ultra-low power devices. We present Hyperdrive: a BWN accelerator dramatically reducing the I/O bandwidth exploiting a novel binary-weight streaming approach, which can be used for arbitrarily sized convolutional neural network architecture and input resolution by exploiting the natural scalability of the compute units both at chip-level and system-level by arranging Hyperdrive chips systolically in a 2D mesh while processing the entire feature map together in parallel. Hyperdrive achieves 4.3 TOp/s/W system-level efficiency (i.e., including I/Os)---3.1x higher than state-of-the-art BWN accelerators, even if its core uses resource-intensive FP16 arithmetic for increased robustness.
We propose a simple and easy to implement neural network compression algorithm that achieves results competitive with more complicated state-of-the-art methods. The key idea is to modify the original optimization problem by adding K independent Gaussian priors (corresponding to the k-means objective) over the network parameters to achieve parameter quantization, as well as an L1 penalty to achieve pruning. Unlike many existing quantization-based methods, our method uses hard clustering assignments of network parameters, which adds minimal change or overhead to standard network training. We also demonstrate experimentally that tying neural network parameters provides less gain in generalization performance than changing network architecture and connectivity patterns entirely.
Deep Neural Networks are highly over-parameterized and the size of the neural networks can be reduced significantly after training without any decrease in performance. One can clearly see this phenomenon in a wide range of architectures trained for various problems. Weight/channel pruning, distillation, quantization, matrix factorization are some of the main methods one can use to remove the redundancy to come up with smaller and faster models. This work starts with a short informative chapter, where we motivate the pruning idea and provide the necessary notation. In the second chapter, we compare various saliency scores in the context of parameter pruning. Using the insights obtained from this comparison and stating the problems it brings we motivate why pruning units instead of the individual parameters might be a better idea. We propose some set of definitions to quantify and analyze units that don't learn and create any useful information. We propose an efficient way for detecting dead units and use it to select which units to prune. We get 5x model size reduction through unit-wise pruning on MNIST.
Quantized Neural Networks (QNNs) are often used to improve network efficiency during the inference phase, i.e. after the network has been trained. Extensive research in the field suggests many different quantization schemes. Still, the number of bits required, as well as the best quantization scheme, are yet unknown. Our theoretical analysis suggests that most of the training process is robust to substantial precision reduction, and points to only a few specific operations that require higher precision. Armed with this knowledge, we quantize the model parameters, activations and layer gradients to 8-bit, leaving at a higher precision only the final step in the computation of the weight gradients. Additionally, as QNNs require batch-normalization to be trained at high precision, we introduce Range Batch-Normalization (BN) which has significantly higher tolerance to quantization noise and improved computational complexity. Our simulations show that Range BN is equivalent to the traditional batch norm if a precise scale adjustment, which can be approximated analytically, is applied. To the best of the authors' knowledge, this work is the first to quantize the weights, activations, as well as a substantial volume of the gradients stream, in all layers (including batch normalization) to 8-bit while showing state-of-the-art results over the ImageNet-1K dataset.
The recent framework of compressive statistical learning aims at designing tractable learning algorithms that use only a heavily compressed representation-or sketch-of massive datasets. Compressive K-Means (CKM) is such a method: it estimates the centroids of data clusters from pooled, non-linear, random signatures of the learning examples. While this approach significantly reduces computational time on very large datasets, its digital implementation wastes acquisition resources because the learning examples are compressed only after the sensing stage. The present work generalizes the sketching procedure initially defined in Compressive K-Means to a large class of periodic nonlinearities including hardware-friendly implementations that compressively acquire entire datasets. This idea is exemplified in a Quantized Compressive K-Means procedure, a variant of CKM that leverages 1-bit universal quantization (i.e. retaining the least significant bit of a standard uniform quantizer) as the periodic sketch nonlinearity. Trading for this resource-efficient signature (standard in most acquisition schemes) has almost no impact on the clustering performances, as illustrated by numerical experiments.