Search

With the increasing usage of social networking platforms seen over recent years, there has been an extensive rise in hate speech usage between the users. Hence, Government and social media platforms face lots of responsibility and challenges to control, detect and eliminate massively growing hateful content as early as possible to prevent future criminal acts such as cyber violence and real-life hate crimes. Since Twitter is used globally by people from various backgrounds and nationalities, the platform contains tweets posted in different languages, including code-mixed language, namely Hindi-English. Due to the informal format of tweets with variations in spelling and grammar, hate speech detection is challenging, especially in code-mixed text containing a mixture of different languages. In this paper, we tackle the critical issue of hate speech on social media, with a focus on a mix of English and Hindi-English (code-mixed) text messages (tweets) on Twitter. We perform hate speech classification using the benefits of character-level embedding representations of tweets and Deep Neural Networks (DNN). We built two architectures, namely Convolutional Neural Network (CNN) and a combination of CNN and Long Short-Term Memory (LSTM) algorithms with character-level embedding as an improvement over Elouali et al. (2020)’s work. Both the models were trained using an imbalanced (original) as well as oversampled (balanced) version of the training dataset and were evaluated on the test set. Extensive experimental analysis was performed by tuning the hyperparameters of our models and evaluating their performance in terms of accuracy, efficiency (runtime) and scalability in detecting whether a tweet is hate speech or non-hate. The performance of our proposed models is compared with Elouali et al. (2020)’s model, and it is observed that our method has an improved accuracy and a significantly improved runtime and is scalable. Among our best performing models, CNN-LSTM performed slightly better than CNN with an accuracy of 88.97%.

Structured pruning methods which are capable of delivering a densely pruned network are among the most popular techniques in the realm of neural network pruning, where most methods prune the original network at a filter or layer level. Although such methods may provide immediate compression and acceleration benefits, we argue that the blanket removal of an entire filter or layer may result in undesired accuracy loss. In this paper, we revisit the idea of kernel pruning (to only prune one or several k × k kernels out of a 3D-filter), a heavily overlooked approach under the context of structured pruning. This is because kernel pruning will naturally introduce sparsity to filters within the same convolutional layer — thus, making the remaining network no longer dense. We address this problem by proposing a versatile grouped pruning framework where we first cluster filters from each convolutional layer into equal-sized groups, prune the grouped kernels we deem unimportant from each filter group, then permute the remaining filters to form a densely grouped convolutional architecture (which also enables the parallel computing capability) for fine-tuning. Specifically, we consult empirical findings from a series of literature regarding Lottery Ticket Hypothesis to determine the optimal clustering scheme per layer, and develop a simple yet cost-efficient greedy approximation algorithm to determine which group kernels to keep within each filter group. Extensive experiments also demonstrate our method often outperforms comparable SOTA methods with lesser data augmentation needed, smaller fine-tuning budget required, and sometimes even much simpler procedure executed (e.g., one-shot v. iterative).

Current communication and network systems are not built for delay-sensitive applications. The most obvious fact is that the communication capacity is only achievable in theory with infinitely long codes, which means infinitely long delays. One remedy for this is to use shorter codes. Conceptually, there is a deeper reason for the difficulties in such solutions: in Shannon's original 1948 paper, he started out by stating that the "semantic aspects" of information is "irrelevant" to communications. Hence, in Shannon's communication system, as well as every network built after him, we put all information into a uniform bit-stream, regardless what meanings they carry, and we transmit these bits over the network as a single type of commodity. Consequently, the network system can only provide a uniform level of error protection and latency control to all these bits. We argue that such a single measure of latency, or Age of Information (AoI), is insufficient for military Internet of Things (IoT) applications that inherently connect the communication network with a cyber-physical system. For example, a self-driving military vehicle might send to the controller a front-view image. Clearly, not everything in the image is equally important for the purpose of steering the vehicle: an approaching vehicle is a much more urgent piece of information than a tree in the background. Similar examples can be seen for other military IoT devices, such as drones and sensors. In this work, we present a new approach that inherently extracts the most critical information in a Military Battlefield IoT scenario by using a metric - called H-Score. This ensures the neural network to only concentrate on the most important information and ignore all background information. We then carry out extensive evaluation of this a by testing it against various inputs, ranging from a vector of numbers to a 1000x1000 pixel image. Next, we introduce the concept of Manual Marginalization, which helps us to make independent decisions for each object in the image. We also develop a video game that captures the essence of a military battlefield scenario and test our developed algorithm here. Finally, we apply our approach on a simple Atari Space Invaders video game to shoot down enemies before they fire at us.

While many computer vision researchers race to architect improved convolutional neural networks (CNNs) and increase task accuracy, relatively little work has been done to understand and quantify the types of errors networks make. Regarding classification networks, certain incorrect responses may be better than others. In an animal classification task, for instance, the network categorizes an input image as one out of a set of trained labels. When such a network gets the label incorrect, there are many wrong labels it can choose. There are, however, varying degrees of incorrectness; a network which classifies a German shepherd as a poodle is a better network than one which classifies the German shepherd as a blue whale. In this work, I explore the responses of CNNs to sets of images in both known classes and novel classes (those which the network was not trained on). I analyze the predicted labels for these sets of images, as well as the predicted label distribution over many images in a given class. This paper also includes a discussion of the hierarchy of predicted labels and true labels, and what a network’s response in terms of higher-level categories reveals about its ability to generalize. Finally, I will discuss how humans and networks each measure the similarity between images, and show novelty's effect on networks' agreement with people.

Background: Diabetic retinopathy is a disease caused due by complications of diabetes mellitus which can lead to blindness. About 33% of the US population with diabetes also show symptoms for diabetes retinopathy. If not treated, diabetic retinopathy worsens over time by progressing through two main pathological stages of non-proliferative and proliferative and four clinical stages. While the diagnostic accuracy of detecting diabetic retinopathy through machine learning have shown to be successful for OCT images, the accuracy of ultra-widefield fundus images have yet to be fully reported. This paper describes a method to non-invasively detect and diagnose diabetic retinopathy from ultra-widefield fundus images. Methods: A total of 62 graded-images were obtained from the Cleveland Clinic. A deep learning algorithm was developed to identify and extract features from the images. The algorithm was then simulated to classify the test images into one of three clinical classes. Data was collected on the accuracy and probability of the diagnosis/classification. Results: The classification algorithm had an average accuracy that ranged from 92% to 97% for the training images and 50% for the test images. Confusion matrices were created to obtain statistical measures of performance such as sensitivity, false negative rate, precision, and the false discovery rate. The sensitivity decreased from 70% to 50% as the image size increased. The precision also decreased from 65% to 50% as the image size increased. Validation methods such as image normalization and transfer learning showed no improvement in classification accuracy. Conclusion: This study demonstrates the potential for applying deep learning algorithms to classify ultra-widefield images. This study also demonstrates the need for doctors to further examine the diagnosis to account for false positives and/or misdiagnosis. Additionally, limitations and their impact on the simulation of the deep learning algorithm were explored.