Modern Convolutional Neural Networks Research Articles

An attention-based deep network for plant disease classification

Plant disease classification using machine learning in a real agricultural field environment is a difficult task. Often, an automated plant disease diagnosis method might fail to capture and interpret discriminatory information due to small variations among leaf sub-categories. Yet, modern Convolutional Neural Networks (CNNs) have achieved decent success in discriminating various plant diseases using leave images. A few existing methods have applied additional pre-processing modules or sub-networks to tackle this challenge. Sometimes, the feature maps ignore partial information for holistic description by part-mining. A deep CNN that emphasizes integration of partial descriptiveness of leaf regions is proposed in this work. The efficacious attention mechanism is integrated with high-level feature map of a base CNN for enhancing feature representation. The proposed method focuses on important diseased areas in leaves, and employs an attention weighting scheme for utilizing useful neighborhood information. The proposed Attention-based network for Plant Disease Classification (APDC) method has achieved state-of-the-art performances on four public plant datasets containing visual/thermal images. The best top-1 accuracies attained by the proposed APDC are: PlantPathology 97.74%, PaddyCrop 99.62%, PaddyDoctor 99.65%, and PlantVillage 99.97%. These results justify the suitability of proposed method.

PP-NAS: Searching for Plug-and-Play Blocks on Convolutional Neural Networks.

Multiscale features are of great importance in modern convolutional neural networks, showing consistent performance gains on numerous vision tasks. Therefore, many plug-and-play blocks are introduced to upgrade existing convolutional neural networks for stronger multiscale representation ability. However, the design of plug-and-play blocks is getting more and more complex, and these manually designed blocks are not optimal. In this work, we propose PP-NAS to develop plug-and-play blocks based on neural architecture search (NAS). Specifically, we design a new search space PPConv and develop a search algorithm consisting of one-level optimization, zero-one loss, and connection existence loss. PP-NAS minimizes the optimization gap between super-net and subarchitectures and can achieve good performance even without retraining. Extensive experiments on image classification, object detection, and semantic segmentation verify the superiority of PP-NAS over state-of-the-art CNNs (e.g., ResNet, ResNeXt, and Res2Net). Our code is available at https://github.com/ainieli/PP-NAS.

EcoFlow: Efficient Convolutional Dataflows on Low-Power Neural Network Accelerators

Dilated and transposed convolutions are widely used in modern convolutional neural networks (CNNs). These kernels are used extensively during CNN training and inference of applications such as image segmentation and high-resolution image generation. We find that commonly-used low-power CNN inference accelerators are <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">not</i> optimized for both these convolutional kernels. Dilated and transposed convolutions introduce significant zero padding when mapped to the underlying spatial architecture, significantly degrading performance and energy efficiency. Existing approaches that address this issue require significant design changes to the otherwise simple, efficient, and well-adopted architectures used to compute direct convolutions. To address this challenge, we propose EcoFlow, a new set of dataflows and mapping algorithms for dilated and transposed convolutions. These algorithms are tailored to execute efficiently on existing low-cost, small-scale spatial architectures and requires minimal changes to existing accelerators. At its core, EcoFlow eliminates zero padding through careful dataflow orchestration and data mapping tailored to the spatial architecture. We evaluate EcoFlow on CNN training workloads and Generative Adversarial Network (GAN) workloads. Experiments in our new cycle-accurate simulator show that, using a common CNN inference accelerator, EcoFlow 1) reduces end-to-end CNN training time between 7-85%, and 2) improves end-to-end GAN training performance between 29-42%, compared to state-of-the-art CNN dataflows.

A Smart Visual Sensor for Smoke Detection Based on Deep Neural Networks.

The automatic detection of smoke by analyzing the video stream acquired by traditional surveillance cameras is becoming a more and more interesting problem for the scientific community thanks to the necessity to prevent fires at the very early stages. The adoption of a smart visual sensor, namely a computer vision algorithm running in real time, allows one to overcome the limitations of standard physical sensors. Nevertheless, this is a very challenging problem, due to the strong similarity of the smoke with other environmental elements like clouds, fog and dust. In addition to this challenge, data available for training deep neural networks is limited and not fully representative of real environments. Within this context, in this paper we propose a new method for smoke detection based on the combination of motion and appearance analysis with a modern convolutional neural network (CNN). Moreover, we propose a new dataset, called the MIVIA Smoke Detection Dataset (MIVIA-SDD), publicly available for research purposes; it consists of 129 videos covering about 28 h of recordings. The proposed hybrid method, trained and evaluated on the proposed dataset, demonstrated to be very effective by achieving a 94% smoke recognition rate and, at the same time, a substantially lower false positive rate if compared with fully deep learning-based approaches (14% vs. 100%). Therefore, the proposed combination of motion and appearance analysis with deep learning CNNs can be further investigated to improve the precision of fire detection approaches.

#176 System to detect vascular calcification through convolutional neural networks

Abstract Background and Aims Cardiovascular disease is the major cause of death in patients with chronic kidney disease (CKD). It is increasingly believed that the presence of vascular calcification (VC) increases the risk of mortality due to cardiovascular events. VC is very common in CKD, especially in hemodialysis patient. To improve the life expectancy of patients, it is important to detect VC in the early stages of the disease. There are several ways to discover VC such as computed tomography (CT) scans, echocardiography, and X-Ray among others. However, one of the best options is the usage of X-ray images because is an affordable and simpler process in comparison to using CT scans or echocardiography. However, the problem with X-ray images is the difficulty of finding a well-trained professional that can detect the VC after analyzing the images which leads to possible misleading diagnoses. The goal of this study is to develop a prototype that detects Vascular Calcification from X-ray images in an automated way using the power of Convolutional Neural Networks (CNN). Method To train a CNN to detect VC from lumbosacral spine X-ray images, a set of 700 DICOM files complemented with 236 JPEG images was available. After getting the images, they had to be anonymized, and data, such as name and age was removed from each image. Then, the images were divided into three datasets: Train, Test, and Validation. Given that the dataset was not balanced between the pictures with VC (Positive) and those with no VC (Negative), it had to be reduced (Table 1). The next step was implemented to clean the images and prepare them to feed the CNN. An example of preprocessing is that some X-ray images were of the entire body. For the analysis, the VC detection was assessed at the anterior and the posterior walls of the abdominal aorta adjacent to vertebrae L1-L4. In addition to that, the DICOM images were converted to JPEG. After having the dataset prepared, a CNN architecture had to be selected. To achieve it, a revision of modern CNN architectures was performed. Two items were considered when selecting these: the performance on the ImageNet Large Scale Visual Recognition Challenge (ILSVRC) and if there was an available pre-trained model. With that criteria, pre-trained EfficientNet-B0 and ResNet50 models were used to build the prototype due to the limited dataset. Results After having the dataset prepared, the architectures selected, and the pre-trained models, the Training, Validation and Test stages were executed. For Training, with EfficientNet-B0, the Accuracy was 86.9%, while for Validation, the Accuracy was 57.4% (Fig. 1). Using the RestNet50 model, the precision was 90.506% during Training, while for Validation, it was 54.8% (Fig. 2). Finally, during the Testing stage, ResNet50 had a better performance, because EfficientNetB0 detected 80% for the negative class and 60% for the positive class. After the CNN models were trained, an additional validation stage was performed to compare the model results with the analysis made by a radiologist and a nephrologist. This comparison determined that the model performed better with images having VC, while it was less accurate with images having no presence of VC. This contrast could have different reasons, one being possible false-positive images. Conclusion This study created a CNN model using pre-trained models on top of modern architectures such as EfficientNet-B0 and ResNet50. Tools were used to prepare the images as well as to train, validate and test the model. The Transfer Learning helped to build a model faster, giving high Accuracy rates with only a few images due to the limited dataset. However, the model could give better results if a wider and balanced set of images is available. Also, a better pre-processing stage can be executed to reduce false-positive images. For future work, this model can be used as the core of a system that not only accurately detects Vascular Calcification from X-Ray images, but also can determine the level of calcification using measures such as the Kauppila Index or Adragao.

Exploring the influence of attention for whole-image mammogram classification

Attention is an important component of modern Convolutional Neural Networks (CNNs) that has been shown to improve baseline model performance for a wide variety of tasks. Attention has shown specific promise in the classification and segmentation of mammograms, but we have a limited understanding of why attention improves performance in these domains. In this paper, we present a robust comparison of different combinations of baseline models and attention methods at two resolutions for whole mammogram classification of masses and calcifications. We find that attention generally helps to improve baseline model performance. However, the extent of improvement is governed by a combination of model architecture and the statistical characteristics of the data. Specifically, we show that high amounts of pooling and model complexity may result in decreased performance for data with high variability. To better understand the effect of attention on mammogram classification, we used LayerCAM, a hierarchical Class Activation Map (CAM) approach, to visualize where the network pays attention in the input image. This research provides statistical evidence that attention can improve the correlation between model performance and LayerCAM activation in the region of interest (ROI). However, these correlations are weak and variable, indicating that improvements in model performance due to attention are not necessarily caused by increased model activation near the ROI. Overall, our work provides novel insights to help guide future efforts in incorporating attention-based mechanisms for mammogram classification.

SAR deep learning sea ice retrieval trained with airborne laser scanner measurements from the MOSAiC expedition

Abstract. Automated sea ice charting from synthetic aperture radar (SAR) has been researched for more than a decade, and we are still not close to unlocking the full potential of automated solutions in terms of resolution and accuracy. The central complications arise from ground truth data not being readily available in the polar regions. In this paper, we build a data set from 20 near-coincident x-band SAR acquisitions and as many airborne laser scanner (ALS) measurements from the Multidisciplinary drifting Observatory for the Study of Arctic Climate (MOSAiC), between October and May. This data set is then used to assess the accuracy and robustness of five machine-learning-based approaches by deriving classes from the freeboard, surface roughness (standard deviation at 0.5 m correlation length) and reflectance. It is shown that there is only a weak correlation between the radar backscatter and the sea ice topography. Accuracies between 44 % and 66 % and robustness between 71 % and 83 % give a realistic insight into the performance of modern convolutional neural network architectures across a range of ice conditions over 8 months. It also marks the first time algorithms have been trained entirely with labels from coincident measurements, allowing for a probabilistic class retrieval. The results show that segmentation models able to learn from the class distribution perform significantly better than pixel-wise classification approaches by nearly 20 % accuracy on average.

Make RepVGG Greater Again: A Quantization-Aware Approach

The tradeoff between performance and inference speed is critical for practical applications. Architecture reparameterization obtains better tradeoffs and it is becoming an increasingly popular ingredient in modern convolutional neural networks. Nonetheless, its quantization performance is usually too poor to deploy (e.g. more than 20% top-1 accuracy drop on ImageNet) when INT8 inference is desired. In this paper, we dive into the underlying mechanism of this failure, where the original design inevitably enlarges quantization error. We propose a simple, robust, and effective remedy to have a quantization-friendly structure that also enjoys reparameterization benefits. Our method greatly bridges the gap between INT8 and FP32 accuracy for RepVGG. Without bells and whistles, the top-1 accuracy drop on ImageNet is reduced within 2% by standard post-training quantization. Extensive experiments on detection and semantic segmentation tasks verify its generalization.

REPrune: Channel Pruning via Kernel Representative Selection

Channel pruning is widely accepted to accelerate modern convolutional neural networks (CNNs). The resulting pruned model benefits from its immediate deployment on general-purpose software and hardware resources. However, its large pruning granularity, specifically at the unit of a convolution filter, often leads to undesirable accuracy drops due to the inflexibility of deciding how and where to introduce sparsity to the CNNs. In this paper, we propose REPrune, a novel channel pruning technique that emulates kernel pruning, fully exploiting the finer but structured granularity. REPrune identifies similar kernels within each channel using agglomerative clustering. Then, it selects filters that maximize the incorporation of kernel representatives while optimizing the maximum cluster coverage problem. By integrating with a simultaneous training-pruning paradigm, REPrune promotes efficient, progressive pruning throughout training CNNs, avoiding the conventional train-prune-finetune sequence. Experimental results highlight that REPrune performs better in computer vision tasks than existing methods, effectively achieving a balance between acceleration ratio and performance retention.

Stable Low-Rank CP Decomposition for Compression of Convolutional Neural Networks Based on Sensitivity

Modern convolutional neural networks (CNNs) play a crucial role in computer vision applications. The intricacy of the application scenarios and the growing dataset both significantly raise the complexity of CNNs. As a result, they are often overparameterized and have significant computational costs. One potential solution for optimizing and compressing the CNNs is to replace convolutional layers with low-rank tensor decomposition. The most suitable technique for this is Canonical Polyadic (CP) decomposition. However, there are two primary issues with CP decomposition that lead to a significant loss in accuracy. Firstly, the selection of tensor ranks for CP decomposition is an unsolved issue. Secondly, degeneracy and instability are common problems in the CP decomposition of contractional tensors, which makes fine-tuning the compressed model difficult. In this study, a novel approach was proposed for compressing CNNs by using CP decomposition. The first step involves using the sensitivity of convolutional layers to determine the tensor ranks for CP decomposition effectively. Subsequently, to address the degeneracy issue and enhance the stability of the CP decomposition, two novel techniques were incorporated: optimization with sensitivity constraints and iterative fine-tuning based on sensitivity order. Finally, the proposed method was examined on common CNN structures for image classification tasks and demonstrated that it provides stable performance and significantly fewer reductions in classification accuracy.

Brain tumor grade classification using the ConvNext architecture.

Brain tumor grade is an important aspect of brain tumor diagnosis and helps to plan for treatment. Traditional methods of diagnosis, including biopsy and manual examination of medical images, are either invasive or may result in inaccurate diagnoses. This study proposes a brain tumor grade classification technique using a modern convolutional neural network (CNN) architecture called ConvNext that inputs magnetic resonance imaging (MRI) data. Deep learning-based techniques are replacing invasive procedures for consistent, accurate, and non-invasive diagnosis of brain tumors. A well-known challenge of using deep learning architectures in medical imaging is data scarcity. Modern-day architectures have huge trainable parameters and require massive datasets to achieve the desired accuracy and avoid overfitting. Therefore, transfer learning is popular among researchers using medical imaging data. Recently, transformer-based architectures have surpassed CNNs for image data. However, recently proposed CNNs have achieved superior accuracy by introducing some tweaks inspired by vision transformers. This study proposed a technique to extract features from the ConvNext architecture and feed these features to a fully connected neural network for final classification. The proposed study achieved state-of-the-art performance on the BraTS 2019 dataset using pre-trained ConvNext. The best accuracy of 99.5% was achieved when three MRI sequences were input as three channels of the pre-trained CNN. The study demonstrated the efficacy of the representations learned by a modern CNN architecture, which has a higher inductive bias for the image data than vision transformers for brain tumor grade classification.

Performance analysis of hybrid deep learning framework using a vision transformer and convolutional neural network for handwritten digit recognition

Digitization created a demand for highly efficient handwritten document recognition systems. A handwritten document consists of digits, text, symbols, diagrams, etc. Digits are an essential element of handwritten documents. Accurate recognition of handwritten digits is vital for effective communication and data analysis. Various researchers have attempted to address this issue with modern convolutional neural network (CNN) techniques. Even after training, CNN filter weights remain unchanged despite the high identification accuracy. As a result, the process cannot flexibly adapt to input changes. Hence computer vision researchers have recently become interested in Vision Transformers (ViTs) and Multilayer Perceptrons (MLPs). The shortcomings of CNNs gave rise to a hybrid model revolution that combines the best elements of the two fields. This paper analyzes how the hybrid convolutional ViT model affects the ability to recognize handwritten digits. Also, the real-time data contains noise, distortions, and varying writing styles. Hence, cleaned and uncleaned handwritten digit images are used for evaluation in this paper. The accuracy of the proposed method is compared with the state-of-the-art techniques, and the result shows that the proposed model achieves the highest recognition accuracy. Also, the probable solutions for recognizing other aspects of handwritten documents are discussed in this paper.•Analyzed the effect of convolutional vision transformer on cleaned and real-time handwritten digit images.•The model's performance improved with the implication of cross-validation and hyper-parameter tuning.•The results show that the proposed model is robust, feasible, and effective on cleaned and uncleaned handwritten digits.

An Adaptive Spatio-temporal Global Sampling for Presentation Attack Detection

Without developing dedicated countermeasures, facial biometric systems can be spoofed with printed photos, replay attacks, silicone masks, or even a 3D mask of a targeted person. Thus, the threat of presentation attacks needs to be addressed to strengthen the security of the biometric systems. Since a 2D convolutional neural network (CNN) captures static features from video frames, the camera motion might hinders the performance of modern CNNs for video-based presentation attack detection (PAD). Inspired by the egomotion theory, we introduce an adaptive spatiotemporal global sampling (ASGS) technique to compensate the camera motion and use the resulting estimation to encode the appearance and dynamics of the video sequences into a single RGB image. This is achieved by adaptively splitting the video into small segments and capturing their global motion within each segment. The proposed global motion is estimated based on four key steps: dense sampling, FREAK feature extraction and matching, similarity transformation, and aggregation function. This allows using deep models pre-trained on images for video-based PAD detection. Moreover, the interpretation of ASGS reveals that the most important parts for supporting the decision on PAD are consistent with motion cues associated with the artifacts, i.e., hand movement, material reflection, and expression changes. Extensive experiments on four standard face PAD databases demonstrate its effectiveness and encourage further study in this domain.

Multi-style spatial attention module for cortical cataract classification in AS-OCT image with supervised contrastive learning

Background and Objective: Precise cortical cataract (CC) classification plays a significant role in early cataract intervention and surgery. Anterior segment optical coherence tomography (AS-OCT) images have shown excellent potential in cataract diagnosis. However, due to the complex opacity distributions of CC, automatic AS-OCT-based CC classification has been rarely studied. In this paper, we aim to explore the opacity distribution characteristics of CC as clinical priori to enhance the representational capability of deep convolutional neural networks (CNNs) in CC classification tasks.Methods: We propose a novel architectural unit, Multi-style Spatial Attention module (MSSA), which recalibrates intermediate feature maps by exploiting diverse clinical contexts. MSSA first extracts the clinical style context features with Group-wise Style Pooling (GSP), then refines the clinical style context features with Local Transform (LT), and finally executes group-wise feature map recalibration via Style Feature Recalibration (SFR). MSSA can be easily integrated into modern CNNs with negligible overhead.Results: The extensive experiments on a CASIA2 AS-OCT dataset and two public ophthalmic datasets demonstrate the superiority of MSSA over state-of-the-art attention methods. The visualization analysis and ablation study are conducted to improve the explainability of MSSA in the decision-making process.Conclusions: Our proposed MSSANet utilized the opacity distribution characteristics of CC to enhance the representational power and explainability of deep convolutional neural network (CNN) and improve the CC classification performance. Our proposed method has the potential in the early clinical CC diagnosis.

FPGA Design of Transposed Convolutions for Deep Learning Using High-Level Synthesis

Deep Learning (DL) is pervasive across a wide variety of domains. Convolutional Neural Networks (CNNs) are often used for image processing DL applications. Modern CNN models are growing to meet the needs of more sophisticated tasks, e.g. using Transposed Convolutions (TCONVs) for image decompression and image generation. Such state-of-the-art DL models often target GPU-based high-performance architectures, due to the high computational and hardware resource needs of TCONV layers. To avoid prohibitive GPU energy costs, CNNs are increasingly deployed to decentralized embedded autonomous devices, such as Field Programmable Gate Arrays (FPGAs). However, this poses challenges for designing efficient hardware implementations of TCONV layers. This paper presents a parameterized design and implementation of a new TCONV module, which is synthesizable onto FPGAs. It is implemented using the High-Level Synthesis (HLS), through a C++ template to parameterize its functional and non-functional properties. These parameters allow kernel sizes, image sizes, quantization and parallelism to be varied by users. With a systematic exploration in this design space, we find an optimal instance of this TCONV module that achieves 6.25 Giga Outputs per Second (Gout/s) using just 1.53 W of power. We then use our TCONV layer in two neural networks for image decompression and image generation. Image decompression achieves a speed throughput of more than 30K frames-per-second (fps) using only the 16% of resources on average, image generation achieves an energy efficiency of 324 fps/W and outperforms comparable state-of-the-art models by at least 7.3×.

Edge Detection via Fusion Difference Convolution.

Edge detection is a crucial step in many computer vision tasks, and in recent years, models based on deep convolutional neural networks (CNNs) have achieved human-level performance in edge detection. However, we have observed that CNN-based methods rely on pre-trained backbone networks and generate edge images with unwanted background details. We propose four new fusion difference convolution (FDC) structures that integrate traditional gradient operators into modern CNNs. At the same time, we have also added a channel spatial attention module (CSAM) and an up-sampling module (US). These structures allow the model to better recognize the semantic and edge information in the images. Our model is trained from scratch on the BIPED dataset without any pre-trained weights and achieves promising results. Moreover, it generalizes well to other datasets without fine-tuning.

Kömür Yakıcılarında Yanma Verimi Tahmini için Gömülü Platformda Çalışabilen Evrişimsel Sinir Ağının Parametre Analizi

Accurately and effectively calculating combustion efficiency in coal burners is crucial for industrial boiler manufacturers. Two main approaches can be used to calculate boiler efficiency: 1) Analyzing the gas emitted from the flue; 2) Visualizing the combustion chamber in the boiler. Flue gas analyzers, which are not user-friendly, come with high costs. Additionally, the physical distance between the flue and the combustion chamber causes the measurement to be delayed. Methods based on visualizing the combustion chamber do not have these disadvantages. This study proposes a system based on visualizing the combustion chamber and has two contributions to the literature: 1) for the first time, the modern Convolutional Neural Networks (CNN) approach is used to estimate combustion efficiency; 2) the CNN architecture with optimal parameters can work on an embedded platform. When classical classification techniques and a CPU-supported processor card are used, efficiency can be calculated from one flame image in 1.7 seconds, while this number increases to approximately 20 frames per second (34 times faster) when the proposed CNN architecture and GPU-supported processor card are used. The results obtained demonstrate the superiority of the proposed CNN architecture and hardware over classical approaches in estimating coal boiler combustion efficiency.

Soil texture prediction with automated deep convolutional neural networks and population-based learning

Convolutional neural networks (CNNs) performance requires tuning of network architectures, which requires machine learning knowledge and significant time and effort. Thus, modern deep CNN in soil spectroscopy faces a major barrier as a result of this process. The components in convolutional neural networks (CNNs) for spectroscopic modelling are usually set heuristically before turning the hyperparameters. This can lead to poorly designed CNN models that are not fully optimised, or to models that take a long time to train to derive the optimal model. Although recent work has shown the importance of tuning the CNN hyperparameters, few studies have sought to tune the CNN components and hyperparameters concurrently. We propose such an approach when designing CNNs for soil spectroscopy, and test its effectiveness with the LUCAS soil library and dataset from the Kellogg Soil Survey Laboratory database. This approach has two phases. First, we automated the process of building the fully connected network which involved automating the selection of the different types of layers and the number of neurons per layer. In the second stage, we automated the selection of the hyperparameters for the network layers, and we experimented with two strategies for selecting the best combination. The first strategy involved adapting the population-based training (PBT) technique by replacing the random search used in PBT with a Bayesian optimisation method, which we call adapted-PBT. In the second strategy, we employed the Bayesian optimisation method. Our study revealed that our adapted-PBT was able to achieve high model performance when compared to using Bayesian optimisation in the second stage when predicting soil texture properties for all three predictive measures used (model efficiency coefficient (MCE), Root Mean Squared Error (RMSE), and the ratio of performance to interquartile range (RPIQ)). The results from our modelling were compared with two recent CNN studies that had used the LUCAS dataset and evaluated using the MCE, RMSE and RPIQ. This comparison indicated improvements (5% to 37%) in performance for all three soil properties (sand, silt and clay) in all performance metrics. These findings provide new insights that could help advance the use of CNNs in soil spectroscopy modelling by concurrently combining the building of the CNN components with tuning the hyperparameters.

Design of a Machine Learning System to Predict the Thickness of a Melanoma Lesion in a Non-Invasive Way from Dermoscopic Images.

Melanoma is the deadliest form of skin cancer, but it can be fully cured through early detection and treatment in 99% of cases. Our aim was to develop a non-invasive machine learning system that can predict the thickness of a melanoma lesion, which is a proxy for tumor progression, through dermoscopic images. This method can serve as a valuable tool in identifying urgent cases for treatment. A modern convolutional neural network architecture (EfficientNet) was used to construct a model capable of classifying dermoscopic images of melanoma lesions into three distinct categories based on thickness. We incorporated techniques to reduce the impact of an imbalanced training dataset, enhanced the generalization capacity of the model through image augmentation, and utilized five-fold cross-validation to produce more reliable metrics. Our method achieved 71% balanced accuracy for three-way classification when trained on a small public dataset of 247 melanoma images. We also presented performance projections for larger training datasets. Our model represents a new state-of-the-art method for classifying melanoma thicknesses. Performance can be further optimized by expanding training datasets and utilizing model ensembles. We have shown that earlier claims of higher performance were mistaken due to data leakage during the evaluation process.

Synapse Compression for Event-Based Convolutional-Neural-Network Accelerators

Manufacturing-viable neuromorphic chips require novel compute architectures to achieve the massively parallel and efficient information processing the brain supports so effortlessly. The most promising architectures for that are spiking/event-based, which enables massive parallelism at low complexity. However, the large memory requirements for synaptic connectivity are a showstopper for the execution of modern convolutional neural networks (CNNs) on massively parallel, event-based architectures. The present work overcomes this roadblock by contributing a lightweight hardware scheme to compress the synaptic memory requirements by several thousand times—enabling the execution of complex CNNs on a single chip of small form factor. A silicon implementation in a 12-nm technology shows that the technique achieves a total memory-footprint reduction of up to <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$374 \times$</tex-math></inline-formula> compared to the best previously published technique at a negligible area overhead.