BISWSRBS: A Winograd-based CNN Accelerator with a Fine-grained Regular Sparsity Pattern and Mixed Precision Quantization

Field-programmable Gate Array (FPGA) is a high-performance computing platform for Convolution Neural Networks (CNNs) inference. Winograd algorithm, weight pruning, and quantization are widely adopted to reduce the storage and arithmetic overhead of CNNs on FPGAs. Recent studies strive to prune the weights in the Winograd domain, however, resulting in irregular sparse patterns and leading to low parallelism and reduced utilization of resources. Besides, there are few works to discuss a suitable quantization scheme for Winograd. In this article, we propose a regular sparse pruning pattern in the Winograd-based CNN, namely, Sub-row-balanced Sparsity (SRBS) pattern, to overcome the challenge of the irregular sparse pattern. Then, we develop a two-step hardware co-optimization approach to improve the model accuracy using the SRBS pattern. Based on the pruned model, we implement a mixed precision quantization to further reduce the computational complexity of bit operations. Finally, we design an FPGA accelerator that takes both the advantage of the SRBS pattern to eliminate low-parallelism computation and the irregular memory accesses, as well as the mixed precision quantization to get a layer-wise bit width. Experimental results on VGG16/VGG-nagadomi with CIFAR-10 and ResNet-18/34/50 with ImageNet show up to 11.8×/8.67× and 8.17×/8.31×/10.6× speedup, 12.74×/9.19× and 8.75×/8.81×/11.1× energy efficiency improvement, respectively, compared with the state-of-the-art dense Winograd accelerator [20] with negligible loss of model accuracy. We also show that our design has 4.11× speedup compared with the state-of-the-art sparse Winograd accelerator [19] on VGG16.

The Influence of Aerial Hyperspectral Image Processing Workflow on Nitrogen Uptake Prediction Accuracy in Maize

A meticulous image processing workflow is oftentimes required to derive quality image data from high-resolution, unmanned aerial systems. There are many subjective decisions to be made during image processing, but the effects of those decisions on prediction model accuracy have never been reported. This study introduced a framework for quantifying the effects of image processing methods on model accuracy. A demonstration of this framework was performed using high-resolution hyperspectral imagery (<10 cm pixel size) for predicting maize nitrogen uptake in the early to mid-vegetative developmental stages (V6–V14). Two supervised regression learning estimators (Lasso and partial least squares) were trained to make predictions from hyperspectral imagery. Data for this use case were collected from three experiments over two years (2018–2019) in southern Minnesota, USA (four site-years). The image processing steps that were evaluated include (i) reflectance conversion, (ii) cropping, (iii) spectral clipping, (iv) spectral smoothing, (v) binning, and (vi) segmentation. In total, 648 image processing workflow scenarios were evaluated, and results were analyzed to understand the influence of each image processing step on the cross-validated root mean squared error (RMSE) of the estimators. A sensitivity analysis revealed that the segmentation step was the most influential image processing step on the final estimator error. Across all workflow scenarios, the RMSE of predicted nitrogen uptake ranged from 14.3 to 19.8 kg ha−1 (relative RMSE ranged from 26.5% to 36.5%), a 38.5% increase in error from the lowest to the highest error workflow scenario. The framework introduced demonstrates the sensitivity and extent to which image processing affects prediction accuracy. It allows remote sensing analysts to improve model performance while providing data-driven justification to improve the reproducibility and objectivity of their work, similar to the benefits of hyperparameter tuning in machine learning applications.

Towards Understanding the Risks of Gradient Inversion in Federated Learning

Federated learning (FL) allows the collaborative training of AI models without needing to share raw data. This capability makes it especially interesting for healthcare applications where patient and data privacy is of utmost concern. However, recent works on the inversion of deep neural networks from model gradients raised concerns about the security of FL in preventing the leakage of training data. In this work, we show that these attacks presented in the literature are impractical in real FL use-cases and provide a new baseline attack that works for more realistic scenarios where the clients’ training involves updating the Batch Normalization (BN) statistics. Furthermore, we present new ways to measure and visualize potential data leakage in FL. Our work is a step towards establishing reproducible methods of measuring data leakage in FL and could help determine the optimal tradeoffs between privacy-preserving techniques, such as differential privacy, and model accuracy based on quantifiable metrics.

Towards Understanding the Risks of Gradient Inversion in Federated Learning

Federated learning (FL) allows the collaborative training of AI models without needing to share raw data. This capability makes it especially interesting for healthcare applications where patient and data privacy is of utmost concern. However, recent works on the inversion of deep neural networks from model gradients raised concerns about the security of FL in preventing the leakage of training data. In this work, we show that these attacks presented in the literature are impractical in real FL use-cases and provide a new baseline attack that works for more realistic scenarios where the clients’ training involves updating the Batch Normalization (BN) statistics. Furthermore, we present new ways to measure and visualize potential data leakage in FL. Our work is a step towards establishing reproducible methods of measuring data leakage in FL and could help determine the optimal tradeoffs between privacy-preserving techniques, such as differential privacy, and model accuracy based on quantifiable metrics.

Comparison of Discriminant Analysis and Adaptive Boosting Classification and Regression Trees on Data with Unbalanced Class

This study aims to determine the best classification results among discriminant analysis, CART, and Adaboost CART on Bank X's Home Ownership Credit (KPR) customers. This study uses secondary data which contains notes on the 5C assessment (Collateral, Character, Capacity, Condition, Capital) and collectibility of current and non-current loans. The sample used in this study was from 2000 debtors. Comparison of classifications based on model accuracy, sensitivity, and overall specificity shows that Adaboost CART is the best method for classifying credit collectibility at Bank X. This is due to the class imbalance in the data. This study compares the classification results between parametric statistics, namely discriminant analysis and non-parametric statistics, namely CART and Adaboost CART. The results of the research can be used as material for consideration and evaluation for banks in determining the policy for providing credit to prospective borrowers from the classification results of KPR Bank X consumers.

Scene terrain classification for autonomous vehicle navigation based on semantic segmentation method

Autonomous transportation is a new paradigm of an Industry 5.0 cyber-physical system that provides a lot of opportunities in smart logistics applications. The safety and reliability of deep learning-driven systems are still a question under research. The safety of an autonomous guided vehicle is dependent on the proper selection of sensors and the transmission of reflex data. Several academics worked on sensor-based difficulties by developing a sensor correction system and fine-tuning algorithms to regulate the system’s efficiency and precision. In this paper, the introduction of vision sensor and its scene terrain classification using a deep learning algorithm is performed with proposed datasets during sensor failure conditions. The proposed classification technique is to identify the mobile robot obstacle and obstacle-free path for smart logistic vehicle application. To analyze the information from the acquired image datasets, the proposed classification algorithm employs segmentation techniques. The analysis of proposed dataset is validated with U-shaped convolutional network (U-Net) architecture and region-based convolutional neural network (Mask R-CNN) architecture model. Based on the results, the selection of 1400 raw image datasets is trained and validated using semantic segmentation classifier models. For various terrain dataset clusters, the Mask R-CNN classifier model method has the highest model accuracy of 93%, that is, 23% higher than the U-Net classifier model algorithm, which has the lowest model accuracy nearly 70%. As a result, the suggested Mask R-CNN technique has a significant potential of being used in autonomous vehicle applications.

Traffic Refinery

Network management often relies on machine learning to make predictions about performance and security from network traffic. Often, the representation of the traffic is as important as the choice of the model. The features that the model relies on, and the representation of those features, ultimately determine model accuracy, as well as where and whether the model can be deployed in practice. Thus, the design and evaluation of these models ultimately requires understanding not only model accuracy but also the systems costs associated with deploying the model in an operational network. Towards this goal, this paper develops a new framework and system that enables a joint evaluation of both the conventional notions of machine learning performance (e.g., model accuracy) and the systems-level costs of different representations of network traffic. We highlight these two dimensions for two practical network management tasks, video streaming quality inference and malware detection, to demonstrate the importance of exploring different representations to find the appropriate operating point. We demonstrate the benefit of exploring a range of representations of network traffic and present Traffic Refinery, a proof-of-concept implementation that both monitors network traffic at 10~Gbps and transforms traffic in real time to produce a variety of feature representations for machine learning. Traffic Refinery both highlights this design space and makes it possible to explore different representations for learning, balancing systems costs related to feature extraction and model training against model accuracy.

Deep Learning-Based Beamforming for Millimeter-Wave Systems Using Parametric ReLU Activation Function

Beamforming design is a crucial stage in millimeter-wave systems with massive antenna arrays. We propose a deep learning network for the design of the precoder and combiner in hybrid architectures. The proposed network employs a parametric rectified linear unit (PReLU) activation function which improves model accuracy with almost no complexity cost compared to other functions. The proposed network accepts practical channel estimation input and can be trained to enhance spectral efficiency considering the hardware limitation of the hybrid design. Simulation shows that the proposed network achieves small performance improvement when compared to the same network with the ReLU activation function.