Regularized Training Framework for Combining Pruning and Quantization to Compress Neural Networks

Forest fires pose a serious threat to the environment with the potential of causing ecological harm, financial losses, and human casualties. While research suggests that climate change will increase the frequency and severity of these fires, recent developments in deep learning and convolutional neural networks (CNN) have greatly enhanced fire detection techniques and capability. These models can be leveraged by unmanned aerial vehicles (UAVs) to automatically monitor burning areas. However, drones can carry only limited computational and power resources; therefore, on-board computing capabilities are constrained by hardware limitations. This work focuses on the design of segmentation models to identify and localize active burning areas from aerial RGB images processed on limited computing resources. To achieve this goal, the research compares the performance of different variants of the DeepLabv3 neural network model for fire segmentation when trained and tested with the FLAME dataset using a k-fold cross validation approach. Experimental results are compared with U-Net, a benchmark model used with the FLAME dataset, by implementing this model in the same codebase as the DeepLabv3 model. This work demonstrates that a refined version of DeepLabv3, with a MobileNetv2 backbone using pretrained layers and a simplified atrous spatial pyramid pooling (ASPP) module, yields a similar performance to U-Net, with a precision of 87.8% and a recall of 83.2%, while only requiring 20% of the number of parameters involved with the U-Net topology. This significantly reduces memory and power consumption, enabling longer UAV flight duration and reducing the processing overhead associated with sensor input, making it more suitable for deployment on unmanned aerial vehicles. The model’s compact architecture, implemented using TensorFlow and Keras for model design and training, along with OpenCV for image preprocessing, makes it portable and easy to integrate with edge devices such as NVIDIA Jetson boards.

Objects and processes classification is a common experimental problem. Its solution, first of all, is needed in automatic diagnosis systems, for example, to determine the equipment operation state through the diagnostic signal or to identify abnormalities with medical images. With the development of convolutional neural networks, new prospects for solving such problems have risen up. However, the classification accuracy that can be achieved on these networks is not sufficient enough for all diagnosis issues. It is subject to, for example, timely diagnosis of the onset of transient phenomena. At the same time, another type of neural network, Kohonen self-organizing maps, has a conceptual property for training on unclassified set of classes, that is, giving the opportunity to solve such an issue. Therefore, the accuracy enhancement of the classification on the basis of Kohonen networks implementation in the architecture of convolutional networks is a relevant objective and has practical significance. The article analyzes the means of improving the accuracy of convolutional neural networks and arising problems solution. The ways of increasing the proportion of correct clustering on Kohonen networks due to the ‘growth’ of its grid shape in determining new classes in the learning process are also given. It is this property that makes it possible to recognize transient phenomena. It is determined that the existing solutions of the combination of Kohonen networks and convolutional networks are aimed at improving the efficiency of only self-organizing maps, with the purpose to improve the accuracy of classification by convolutional networks, it became necessary to develop a new architecture. The paper provides a description of this issue. Since the initial information of the Kohonen networks is the weighting matrix values of the grid shape neurons, it was necessary to associate it with the representation of the images in order to process the diagnostic images. The paper proposes the concept of a built-in associative array block based on Kohonen networks. According to the proposed method, a software implementation of a hybrid neural network is developed. The formulation and results of computational experiments are presented. The efficiency of the proposed method is experimentally proved. Keywords: convolutional neural networks, CNN, Kohonen self-organizing maps, SOM, signal classification, image classification.

The report presents an algorithm for constructing a convolutional hierarchical neural network classifier, which is a modification of the algorithm for constructing hierarchical neural network classifiers suggested before. The original algorithm was designed to exploit intrinsic class hierarchy to build a class tree with a neural network in each node classifying groups of initial classes (in a non-terminal node) or a subset of original classes (in a terminal node). The convolutional modification utilizes convolutional neural networks instead of regular fully connected networks in order to apply the model to image classification tasks. Use of class hierarchy for image classification should reduce the number of adjusted neural network parameters compared to deep convolutional neural networks, and therefore it should reduce training and inference time. In this context the algorithm may be compared with some pruning techniques. The convolutional hierarchical neural network classifier inherits some hyperparameters of a conventional hierarchical neural network classifier, like the activation threshold and the threshold by the share of voting patterns. The goal of this study was to explore different strategies of choosing these hyperparameters. To test these strategies, we used the CIFAR-10 dataset. Also, for demonstration purposes we apply the convolutional hierarchical neural network classifier to the CIFAR-100 dataset.

Hyperspectral Image (HSI) classification is one of the fundamental tasks in the field of remote sensing data analysis. CNN (Convolutional Neural Network) has been proven to be an effective deep learning model, which can extract high-level features directly from the raw data and thereby utilize rich information contained in HSI data. However, labor cost to label enough HIS data for training model is usually expensive, so that it is a strong demand of utilizing limited training data to get a satisfied classification accuracy. In this paper, we put forward a deep cube CNN model – DCNR, which is composed of a cube neighbor HSI pixels strategy, a deep CNN and a random forest classifier. In DCNR model, cubic samples, containing spectral-spatial information, are generated by putting each target pixel and its neighbors together. Then features with high representative ability, extracted by applying a specially designed cube CNN model on each cubic sample, are fed into the random forest classifier for the classification of the target pixel. Results show that DCNR model can achieve classification accuracy of 96.78%, 96.08% and 94.85% on KSC, IP and SA datasets respectively with 20% samples as training set, and 85.03%, 83.45 and 62.17% on KSC, IP and SA datasets respectively with only 1% samples as training set, significantly outperforming random forest and cube CNN models.

Objectives: The paper presents application of convolution neural network and artificial neural network for image classification problem for clothing dataset along with their performance comparison against different optimizers. The major objective of this paper is to perform image classification on fashion-mnist clothing dataset images. Methods: The methods used here are, the traditional ANN and CNN. Here image classification is performed on Fashion-mnist, clothing dataset using CNN and ANN with different optimizers. The performance of the working of ANN and CNN in classifying images from fashion-mnist dataset is compared against different optimizers namely stochastic gradient Descent, Adagrad, RMS prop and Adam optimizer. Findings: The study found that CNN worked better than ANN yielding training accuracy of 95%, 93% and testing accuracy of 91%, 89% when used with Adam and RmsProp respectively. Novelty: The novelty of this work is to present a comparative study of image classification using CNN, ANN using different optimizers, since not many studies or research articles showed the performance comparison of traditional and convolution neural networks in image classification along with different optimizers. Since the real-world scenarios of today require enormous data to be processed, CNN can fit well to diversify applications since they highly reduce the number of parameters to be trained that speeds up the training process. Moreover, to be specific on image classification problems they require the best and most prominent features to be detected and uncovered; this can be achieved using CNN since it has the concept of convolution using filters at its Core. Hence, CNN is highly recommended for such image classification applications than the traditional artificial-neural-networks because of the aforementioned reasons. Keywords: ANN (Artificial NeuralNetworks); CNN (ConvolutionNeural Networks); Optimization; SGD (Stochastic Gradient Descent); RmsProp (Root mean Square propagation); Adam (Adaptive moment estimation); AdaGrad (Adaptive Gradient)

Most popular image classification methods mainly focus on classification ability rather than recognizing new things. However, human lay emphasis on cognition first and then classification, which is closely related to human memory system. Though many memory models have been proposed, they are studied in word list whereas the reports about natural images are still limited. This paper proposes a memory model for image recognition and classification based on convolutional neural network and Bayesian decision. First the image feature is extracted by convolutional neural network and stored in binary form. Then the representation, storage and retrieval processes of visual images are modeled. The test image feature vector is matched in parallel to the studied image vectors, and the likelihood values are calculated. Finally, the odd that the test image belongs to a new class is computed based on all likelihood values. If the odd value is greater than a certain threshold, the test image is regarded as new; otherwise, the Bayesian decision rule for image classification is performed. Experimental results on Caltech-101 and Caltech-256 datasets show that the proposed method can perform well in image recognition and classification tasks. And the hit probability of the method is higher than two typical methods, SRC and ELM, at present while the false alarm rate is far lower than them.

As one of the indispensable components in convolutional neural network (CNN), loss function assists in updating parameters of CNN models during the training phase. Generally, different loss functions can assist convolutional neural network (CNN) to learn different feature representations, and different feature representations can be treated as different knowledge learned from objects. In this paper we introduce a novel training framework, namely Loss Transferring (LT), to improve the generalization ability of CNN. LT contains multiple training phases, and each training phase uses a different loss function. Under this framework, CNN models can combine different knowledge of objects by transferring the knowledge learned via one loss function to another. LT contains two components, i.e., loss function set and training strategy. In order to build appropriate loss function set, we establish two basic guides. And according to these basic guides, we design a new loss function in the last layer of CNN models (layer before softmax operation), namely Near Classifier Hyper-Plane (N-CHP) loss, which makes the learned object features belonging to the same category have the minimum intra-class distance and be near the classifier hyper-plane. Based on the two loss function set $\{\text {MSE, softmax}\}$ and $\{\text {N-CHP, softmax}\}$ , we setup two specific training methods, $LT_{\text {MSE, softmax}}$ and $LT_{\text {N-CHP, softmax}}$ , which can be universally applied to different CNN models with low additional computation cost. Meanwhile, two training strategies, multi-phase strategy 1 and multi-phase strategy 2, are further proposed to improve the training efficiency of LT. Extensive experimental results on shallow, moderate and deep models with four benchmark datasets, including MNIST, SVHN, CIFAR-10 and CIFAR-100, demonstrate that CNN models can bring obvious performance improvements when working with $LT_{\text {MSE, softmax}}$ and $LT_{\text {N-CHP, softmax}}$ , which verifies the effectiveness of LT and the proposed two basic guides.

In recent years Neural Networks and especially Deep Neural Networks (DNN) have seen a rise in popularity. DNNs have increased the overall performance of algorithms in applications such as image recognition and classification, 2D and 3D pose detection, natural language processing, and time series forecasting. Especially in image classification and recognition, so-called Convolutional Neural Network (CNN) have gained high interest as they can reach a high accuracy, which makes them a viable solution in cancer detection or autonomous driving. As CNNs are widely used in image tasks, different visualizations techniques have bee developed to show how their internals are working. Apart from image tasks, CNNs are applicable to other problems, e.g., time series classification, time series forecasting, or natural language processing. CNNs in those contexts behave similarly, allowing for the same visualization techniques to make them more interpretable. In this chapter, we adapt image visualization algorithms to time series problems, allowing us to build granular, intuitively interpretable feature hierarchies to make a time series forecast as understandable as an image recognition task. We do so by using our previous work on power time series forecasting using CNN Auto Encoders (AEs) and applying typical CNN visualization techniques to it. Thus, we guid computer scientists to provide better interpretable figures for a time series forecasting task to application domain experts.KeywordsVisualizationConvolutional neural networksTime series data

A genetic algorithm (GA), is used to optimize the many parameters of a convolutional neural network (CNN) that control the structure of the network. CNNs are used in image classification problems where it is necessary to generate feature descriptors to discern between image classes. Because of the deep representation of image data that CNNs are capable of generating, they are increasingly popular in research and industry applications. With the increasing number of use cases for CNNs, more and more time is being spent to come up with optimal CNN structures for different applications. Where one CNN might succeed at classification, another can fail. As a result, it is desirable to more easily find an optimal CNN structure to increase classification accuracy. In the proposed method, a GA is used to evolve the parameters that influence the structure of a CNN. The GA compares CNNs by training them on images of concrete containing cracks. The best CNN after several generations of the GA is then compared to the state-of-the-art CNN for crack detection. This work shows that it is possible to generalize the process of optimizing a CNN for image classification through the use of a GA.

Deep learning performs a great role in solving many problems, particularly with image classification. Still, the performance of many architectures is often limited at some points, especially with regular architecture. In this paper, we propose an efficient and noncomplex image classification architecture using deep learning based on the most popular algorithms: Convolutional Neural Network (CNN) for Feature extraction, Stacked AutoEncoder (SAE) for reducing the dimensionality, and Recurrent Neural Network (RNN) for increasing the accuracy. The proposed network, called Convolutional Stacked AutoEncoder Recurrent Neural Network (CSAERNet), that takes the input to be features-extracted by CNN then put as input to SAE for dimensionality reduction after that the RNN takes it as input to increase the accuracy. The proposed CSAERNet has been evaluated on two widely-used benchmark datasets; CIFAR-10 and CIFAR-100. The obtained results are compared with the latest state-of-the-art approaches, the results showed that the proposed CSAERNet perform well on both datasets and outperforms on other models.

Convolutional Neural Networks (CNNs) have emerged as a fundamental technology for machine learning. High performance and extreme energy efficiency are critical for deployments of CNNs, especially in mobile platforms such as autonomous vehicles, cameras, and electronic personal assistants. This paper introduces the Sparse CNN (SCNN) accelerator architecture, which improves performance and energy efficiency by exploiting the zero-valued weights that stem from network pruning during training and zero-valued activations that arise from the common ReLU operator. Specifically, SCNN employs a novel dataflow that enables maintaining the sparse weights and activations in a compressed encoding, which eliminates unnecessary data transfers and reduces storage requirements. Furthermore, the SCNN dataflow facilitates efficient delivery of those weights and activations to a multiplier array, where they are extensively reused; product accumulation is performed in a novel accumulator array. On contemporary neural networks, SCNN can improve both performance and energy by a factor of 2.7x and 2.3x, respectively, over a comparably provisioned dense CNN accelerator.

Convolutional neural networks (CNNs) are widely employed in many image recognition applications. With the proliferation of embedded and mobile devices, such applications are becoming commonplace on mobile devices. Network pruning is a commonly used strategy to reduce the memory and storage footprints of CNNs on mobile devices. In this article, we propose customized versions of the sparse matrix multiplication algorithm to speed up inference on mobile devices and make it more energy efficient. Specifically, we propose a Block Compressed Sparse Column algorithm and a bit-representation-based algorithm (BitsGEMM) that exploit sparsity to accelerate the fully connected layers of a network on the NVIDIA Jetson TK1 platform. We evaluate the proposed algorithms using real-world object classification and object detection applications. Experiments show that performance speedups can be achieved over the original baseline implementation using cuBLAS. On object detection CNNs, an average speedup of 1.82× is obtained over baseline cuBLAS in the fully connected layer of the VGG model, whereas on classification CNNs, an average speedup of 1.51× is achieved for the fully connected layer of the pruned-VGG model. Energy consumption reduction of 43--46% is also observed due to decreased computational and memory bandwidth demands.

Convolutional Neural Networks (CNNs) have emerged as a fundamental technology for machine learning. High performance and extreme energy efficiency are critical for deployments of CNNs, especially in mobile platforms such as autonomous vehicles, cameras, and electronic personal assistants. This paper introduces the Sparse CNN (SCNN) accelerator architecture, which improves performance and energy efficiency by exploiting the zero-valued weights that stem from network pruning during training and zero-valued activations that arise from the common ReLU operator. Specifically, SCNN employs a novel dataflow that enables maintaining the sparse weights and activations in a compressed encoding, which eliminates unnecessary data transfers and reduces storage requirements. Furthermore, the SCNN dataflow facilitates efficient delivery of those weights and activations to a multiplier array, where they are extensively reused; product accumulation is performed in a novel accumulator array. On contemporary neural networks, SCNN can improve both performance and energy by a factor of 2.7x and 2.3x, respectively, over a comparably provisioned dense CNN accelerator.

Abstract. Deep learning models have demonstrated remarkable success in fields such as language processing and computer vision, routinely employed for tasks like language translation, image classification, and anomaly detection. Recent advancements in ocean sciences, particularly in data assimilation (DA), suggest that machine learning can emulate dynamical models, replace traditional DA steps to expedite processes, or serve as hybrid surrogate models to enhance forecasts. However, these studies often rely on ocean models of intermediate complexity, which involve significant simplifications that present challenges when transitioning to full-scale operational ocean models. This work explores the application of convolutional neural networks (CNNs) in data assimilation within the context of the HYbrid Coordinate Ocean Model (HYCOM) in the Gulf of Mexico. The CNNs are trained to correct model errors from a 2-year, high-resolution (1/25°) HYCOM dataset, assimilated using the Tendral Statistical Interpolation System (T-SIS). The CNNs are trained to replicate the increments generated by the T-SIS data assimilation package, aiming to correct model forecasts of sea surface temperature (SST) and sea surface height (SSH). The inputs to the CNNs include real satellite observations of SST from the Group for High Resolution Sea Surface Temperature (GHRSST), along-track altimeter SSH observations (ADT), the model background state (previous forecast), and the innovations (differences between observations and background). We assess the performance of the CNNs across five controlled experiments, designed to provide insights into their application in environments governed by full primitive equations, real observations, and complex topographies. The experiments focus on evaluating (1) the architecture and complexity of the CNNs, (2) the type and quantity of observations, (3) the type and number of assimilated fields, (4) the impact of training window size, and (5) the influence of coastal boundaries. Our findings reveal significant correlations between the chosen training window size – a factor not commonly examined – and the CNNs' ability to assimilate observations effectively. We also establish a clear link between the CNNs' architecture and complexity and their overall performance. This research uses artificial intelligence to enhance ocean forecasting in the Gulf of Mexico. By using convolutional neural networks, the study improves predictions of sea temperatures and heights by integrating real satellite data with existing models. Through five comprehensive experiments, the team found that the amount of training data and the design of the neural networks significantly affect accuracy. These insights pave the way for faster, more reliable ocean models, benefiting environmental monitoring and maritime operations.

Power-efficient data processing subsystems performing millions of complex concurrent arithmetic operations per second form part of today’s essential solution required to meet the growing demand of edge computing applications, given the volume of data collected by real-time Internet-Of-Things (IoT) sensors. Adding to it, the in-memory computation designed as memory and processing elements on a single wafer has enabled promising performance improvement in terms of computational power savings by avoiding the memory wall created while accessing the memory array. The Resistive RAM (RRAM), with its simple metal-insulator-metal (MIM) structure, proves to be a very appealing candidate for in-memory computation given its ultralow switching power and its Complementary Metal Oxide Semiconductor (CMOS) process fabrication compatibility. However, despite all advantages, the resistive switching phenomenon in RRAM has an inherent stochastic variability. On the algorithmic side, convolution neural networks (CNN) have gained popularity in image classification applications, and the network’s architecture is memory-intense in nature for memorizing the trained weights. Hence, an RRAM-based CNN system will pave way for a power-efficient image classification system on the edge. Accounting however for the inherent variability in RRAM (inter-device and intra-device), the accuracy of CNN’s prediction is surely expected to drop. This motivates us to quantify the impact of RRAM variability on the CNN trained weights and classification accuracy (prediction loss). In this study, we have constructed a Look-Up-Table (LUT) based model for encoding wide current compliance (2μA to 250μA) 65nm CMOS 1T1R OxRAM’s (TiN/HfO2/Hf/TiN) resistive variability into CNN’s trained weight in a digital regime. The RRAM resistance encoded trained weights are in turn used here to simulate the two extreme CNN architectures, namely, Fully Serial System (FSS) and Fully Parallel System (FPS). The architectures’ prediction variability trends are quantified given its current compliance, RRAM resistive variability, CNN’s convolution matrix sizes (5×5, 3×3, 1×1, and 1×1 max pool), the total number of layers in the CNN as well as the input image pixel size.