Welcome and opening remarks

African languages are spoken by over a billion people, but are underrepresented in NLP research and development. The challenges impeding progress include the limited availability of annotated datasets, as well as a lack of understanding of the settings where current methods are effective. In this paper, we make progress towards solutions for these challenges, focusing on the task of named entity recognition (NER). We create the largest human-annotated NER dataset for 21 African languages, and we study the behaviour of state-of-the-art cross-lingual transfer methods in an Africa-centric setting, demonstrating that the choice of source language significantly affects performance. We show that choosing the best transfer language improves zero-shot F1 scores by an average of 16 points across 21languages compared to using English. Our results highlight the need for benchmark datasets and models that cover typologically-diverse African languages.

Patient-specific features used for surgical planning and custom implant design may require 3D reconstruction of the shape and pose of organs of interest from medical images. Such a task is often accomplished through a challenging problem of fitting a parametric model to the 3D geometry using energy optimization. In this work, we propose a novel generative model-based deep convolutional autoencoder to reconstruct multiple organs in 3D from a 2D image. To this end, we combine a convolutional encoder network with a parametric geometric model that serves as a decoder. The resulting reconstructions compare favorably with current state-of-the-art approaches in terms of accuracy.

Building a usable radio monitoring automatic speech recognition (ASR) system is a challenging task for under-resourced languages and yet this is paramount in societies where radio is the main medium of public communication and discussions. The main challenge is the absence of transcribed radio speech datasets. In this paper, we create a Luganda radio dataset and build a COVID-19 ASR. We use the ASR to analyse public radio discussions for public health response. We openly release a radio speech corpus of 155 hours. To our knowledge, this is the first publicly available radio dataset in sub-Saharan Africa.

Self-supervised contrastive learning is among the recent representation learning methods that have shown performance gains in several downstream tasks including semantic segmentation. This paper evaluates strong data augmentation, one of the most important components for self-supervised contrastive learning's improved performance. Strong data augmentation involves applying the composition of multiple augmentation techniques on images. Surprisingly, we find that the existing data augmentations do not always improve performance for semantic segmentation for medical images. We experiment with other augmentations that provide improved performance.

Adversarial learning represents a powerful technique for generating data statistics. Its successful implementation in quantum computational platforms is not straightforward due to limitations in connectivity, quantum operation fidelity, and limited access to the quantum processor for statistically relevant results. Constraining the number of quantum operations and providing a design with a low compilation cost, we propose a quantum generative adversarial network design that uses real Hilbert spaces as the framework for the generative model. We consider quantum generator and discriminator architectures based on a variational quantum circuit. For low-depth ans\"atze designs, we consider the real Hilbert space as the working space for the quantum adversarial game. This architecture improves state-of-the-art quantum generative adversarial performance while maintaining a shallow-depth quantum circuit and a reduced parameter set. We tested our design in a low resource regime, generating handwritten digits with the MNIST as the reference dataset. We could generate undetected data (digits) with just 15 epochs working in the real Hilbert space of 2, 3, and 4 qubits. Our design uses native quantum operations established in superconducting-based quantum processors and is compatible with ion-trapped-based architectures.

Image and video understanding goal is to make inferences about the surrounding world from the corresponding image or video data, e.g., identification and localization of objects. It can also be extended towards identifying actions and/or recognizing relations between objects. It has attracted great attention in the research community both because of its widespread applications (e.g., to automated driving and robotics) and the fascinating scientific and engineering challenges that it brings (e.g., designing a system that can learn about the 3D, time-vary world from a mere video). Through the use of deep learning, great advances in scene/video understanding have been seen in recent years. A major limitation of most such approaches is that they require large-scale labelled data for learning, where the annotation cost can be expensive, especially when annotating pixel-wise segmentation masks in videos. In this talk, I will focus on how to learn a scene and video understanding provided with a few labelled examples, and to use the interpretability of deep spatiotemporal models to give us insights on how to improve their generalization capabilities. This research direction has the potential to decolonize Computer Vision by enabling developing countries with limited resources and labelled data to contribute to the field and work on applications that serve their own communities.

There may exist multiple models that perform almost equally well for any given prediction task. We examine how predictions change across these competing models. In particular, we study predictive multiplicity -- in probabilistic classification. We formally define measures for our setting and develop optimization-based methods to compute these measures for convex problems. We apply our methodology to gain insight into why predictive multiplicity arises. We demonstrate the incidence and prevalence of predictive multiplicity in real-world risk assessment tasks. Our results emphasize the need to report multiplicity more widely.

End-to-End (E2E) approach to Automatic Speech Recognition (ASR) is a hot research agenda. It is interesting for less-resourced languages (LRL) since it avoids the use of a pronunciation dictionary. However, E2E is data greedy, which makes the application of E2E to LRL questionable. However, using data from other languages in a multilingual (ML) setup is being applied to solve the problem of data scarcity. We have conducted ML E2E ASR experiments for four less-resourced Ethiopian languages using different language and acoustic modeling units. The results of our experiments show that relative Word Error Rate (WER) reductions (over the monolingual E2E systems) of up to 29.83% can be achieved by just using data from two related languages in E2E ASR system training. Moreover, we have also noticed that the use of data from less related languages also leads to E2E ASR performance improvement over the use of monolingual data.

This paper presents empirically driven recommendations for advancing the use of spoken language systems in children's language education. We propose shifts in the current paradigm of machine learning research to better fit the growing needs of educators as well as capture concerns expressed by those in the field of AI.

Announcements.

There may exist multiple models that perform almost equally well for any given prediction task. We examine how predictions change across these competing models. In particular, we study predictive multiplicity -- in probabilistic classification. We formally define measures for our setting and develop optimization-based methods to compute these measures for convex problems. We apply our methodology to gain insight into why predictive multiplicity arises. We demonstrate the incidence and prevalence of predictive multiplicity in real-world risk assessment tasks. Our results emphasize the need to report multiplicity more widely.

Over the last decade, there has been a lot of artificial intelligence (AI)-based solutions proposed in healthcare. However, only a few of the solutions are clinically in use. Lack of trust in healthcare AI-based solutions is tied to the technical characteristics of AI, and how these properties can be understood clinically or biologically. Explainable AI (XAI) can improve the interpretability of AI-based solutions, providing qualitative and quantitative reasons for how AI models make their decisions. In this study, we compare XAI tools: Shapely Addictive exPlanations (SHAP) and Local Interpretable Model-Agnostic Explanations (Lime). Finally, propose linking quantitative imaging features to biology.

In spite of many proposals to solve the neural machine translation problem of low-resource languages, it continues to be difficult for languages with few resources. The issue becomes even more complicated when few resources cover only a single domain. In our attempt to combat this issue, we propose a new approach to improve NMT for low-resource languages. The proposed approach using the transformer model shows 5.3, 5.0, and 3.7 BLEU score improvement for Gamo-English, Gofa-English, and Dawuro-English language pairs, respectively. We discuss our contributions and envisage future steps in this challenging research area.

Patient-specific features used for surgical planning and custom implant design may require 3D reconstruction of the shape and pose of organs of interest from medical images. Such a task is often accomplished through a challenging problem of fitting a parametric model to the 3D geometry using energy optimization. In this work, we propose a novel generative model-based deep convolutional autoencoder to reconstruct multiple organs in 3D from a 2D image. To this end, we combine a convolutional encoder network with a parametric geometric model that serves as a decoder. The resulting reconstructions compare favorably with current state-of-the-art approaches in terms of accuracy.

Irish potato is among the important food and cash crops to most smallholder farmers in Tanzania. Despite its importance in household economy and food security, yields are generally low due to the effects of diseases, specifically Early and Late blight. The current management of these two diseases includes the removal of the affected leaves and plants to reduce their spread, signifying that early detection is the key to successful management. This study therefore developed a Machine Learning model to detect early these two diseases based on leaf imagery data and enable the farmer to make appropriate decision for managing the spread of the diseases. Resnet152 and Inceptionv3 Convolution Neural Network architectures were used to train the model in a dataset of 50,310 imagery samples. The results showed that Resnet152 achieved an accuracy of 83.4% while Inceptionv3 achieved an accuracy of 80.1%. These results demonstrate the suitability of our model to early detect Early and Late blight diseases.

Conventional identification of leukemia based on visual inspection of blood smears through a microscope is time-consuming, error-prone, and is limited by the hematologist’s physical acuity. Therefore, an automated optical image processing system is required to support clinical decision-making. To address this problem, we developed a machine learning-based real-time automated diagnostic system to assist medical care workers. Blood smear slides (n = 250) were prepared from clinical samples, imaged, and analyzed in Jimma Medical Center, Hematology department. The system was able to categorize four common types of leukemia’s through a robust image segmentation protocol, followed by classification using the support vector machine. It was able to classify leukemia types with an accuracy, sensitivity, and specificity of 97.69%, 97.86% and 100%, respectively for the test datasets, and 97.5%, 98.55% and 100%, respectively, for the validation datasets. The computer-assisted diagnosis system took less than one minute for processing and assigning the leukemia types, compared to an average period of 30 minutes by unassisted manual approaches. Moreover, the automated system complements the healthcare workers’ in their efforts, by improving the accuracy rates in diagnosis from ∼70% to over 97%. Importantly, our module is designed to assist the healthcare facilities in the rural areas of sub-Saharan Africa, equipped with fewer experienced medical experts, especially in screening patients for blood associated diseases including leukemia.

This paper presents empirically driven recommendations for advancing the use of spoken language systems in children's language education. We propose shifts in the current paradigm of machine learning research to better fit the growing needs of educators as well as capture concerns expressed by those in the field of AI.

A bulk of real-life documents contain vital information and knowledge about history, culture,economy, politics, religion, and science that are written in Ethiopic script. This knowledge has to be shared andthe advancement of technology like Optical Character Recognition (OCR) brings the need to digitize documentsand make them available for public use. OCR is a process that allows printed, typewritten, and handwritten textto be recognized optically and converted into a machine-readable format that can be accepted by a computer forfurther processing. Nowadays, effective OCR systems have been developed for languages, like English that haswider use internationally. Researches in the area of Amharic OCR are ongoing since 1997. Attempts were madein adopting recognition algorithms to develop Amharic OCR. This study is, thus, an attempt made to develop anOCR system for real-life documents written in Ethiopic characters. In this study we propose a novel featureextraction schema using Gabor Filter and Principal Component Analysis (PCA), followed by a GeneticAlgorithm (GA) based on supported vector machine classifier (SVM). The prototype was tested on real-lifeEthiopic documents such as books, newspapers, and magazines, in which an average accuracy of 98.33% forEthiopic characters is registered.

Word embeddings powered the early days of neural network-based NLP research. Their effectiveness in small data regimes makes them still relevant in low-resource environments. However, they are limited in two critical ways: linearly increasing memory requirements and out-of-vocabulary token handling. In this work, we present a distillation technique of word embeddings into a CNN network using contrastive learning. This method allows embeddings to be regressed given the characters of a token. It is then used as a pretrained layer, replacing word embeddings. Low-resource languages are the primary beneficiary of this method and hence, we show its effectiveness on two morphology-rich Semitic languages, and in a multilingual NER task comprised of 10 African languages. The resulting model is a data efficient one that improves performance, lowers memory requirement and supports transfer of word representation out of the box.

Tuberculosis (TB) is curable, and millions of deaths could be averted if diagnosed early. One of the sources of screening for TB is chest x-rays. Still, its success depends on the interpretation of skilled and experienced radiologists, mostly lacking in high TB burden regions. However, with the intervention of a computer-aided detection system, TB can be automatically detected from chest x-rays. This paper presents an Ensemble model based on multiple pre-trained models to automatically detect TB from chest x-rays. The models were trained on the Shenzhen dataset and validated on the Montgomery dataset to achieve good generalization on a new (unseen) dataset. The proposed Ensemble model achieved high accuracy and sensitivity that is comparable with state-of-the-art models and outperformed existing Ensemble models aimed at Tuberculosis classification.

The recent deep learning breakthroughs in language and vision tasks can be mainly attributed to large-scale transformers. Unfortunately, their massive size and high compute requirements have limited their use in resource-constrained environments. Dynamic neural networks promise reduced amount of compute requirement by dynamically adjusting the computational path based on the input. We propose a layer skipping dynamic vision transformer (DynamicViT) that skips layers for each sample based on decisions given by a reinforcement learning agent. Extensive experiment on CIFAR-10 and CIFAR-100 showed that this dynamic ViT gained an average of 40\% speed increase evaluated on different batch sizes ranging from 1 to 1024.

People’s opinions and judgements regarding a particular movie differ. A movie that is thoroughly enjoyed and recommended by an individual might be hated by another. One characteristic of humans is the ability to have feelings which could be positive or negative. To automatically classify and study human feelings, an aspect of natural language processing, sentiment analysis and opinion mining were designed to understand human feelings regarding several issues which could affect a product, a social media platforms, government, or societal discussions or even movies. Several works on sentiment analysis have been done on high resource languages while low resources languages like Yoruba have been sidelined. Due to the scarcity of datasets and linguistic architectures that will suit low resource languages, African languages "low resource languages" have been not fully explored. For this reason, our attention is placed on Yoruba to explore sentiment analysis on reviews of Nigerian movies. The data comprised 1500 movie reviews that were sourced from IMDB, Rotten Tomatoes, Letterboxd, Cinemapointer and Nollyrated. We develop sentiment classification models using the state-of-the-art pre-trained language models like mBERT and AfriBERTa to classify the movie reviews.

Cognitive and neuroscience of attention researches suggest the use of spatio-temporal features for an efficient video saliency prediction. This is due to the representative nature of spatio-temporal features for data collected across space and time, such as videos. Video saliency prediction aim to find visually salient regions in a stream of images. Many video saliency prediction models are proposed in the past couple of years. However, the problem still remains a considerable challenge. This is mainly due to the complex nature of video saliency prediction and scarcity of representative saliency benchmarks. Given the importance of saliency identification for various computer vision tasks, revising and enhancing the performance of video saliency prediction models is crucial. To this end, we propose a novel interactive video saliency prediction model that employ stacked convLSTM based architecture along with a novel XY-shift frame differencing custom layer. Specifically, we have introduced an encoder-decoder based architecture with a prior layer undertaking XY-shift frame differencing, a residual layer fusing spatially processed (VGG-16 based) features with XY-shift frame differenced frames, and a stacked convLSTM component. Extensive experimental results over the largest video saliency dataset, DHF1K, show the competitive performance of our model against the state-of-the-art models.

The use of automatic speech recognition (ASR) systems for spoken languages has become widespread recently. Contrarily, the vast majority of African languages have limited linguistic resources to sustain the robustness of these systems. We present a study on an end-to-end speech recognition system for Nigerian-Pidgin-English. Using our unique dataset, we fine-tuned different variants of the Wac2Vec2.0 architecture. We contrasted the results of these techniques with those of preceding studies. Empirically, we achieved a low word error rate of 33\% on the test set outperforming the baseline method and also surpassed other variants of the Wac2Vec2.0 architecture in terms of qualitative assessments.

The goal of this study is to develop a deep learning model to predict the early occurrence of Streak Virus and Lethal Necrosis in maize leaves and then deploy the model in a mobile app. We developed a convolutional neural network model from scratch and trained it using 1500 images dataset belonging to three classes namely; HEALTHY, MLN and MSV which attained a validation accuracy of 98.44%. However, we plan to develop pre-trained models to compare the results with those already attained and select the best model to be deployed in a mobile app for early detection and testing in real life environment.

Skin diseases are highly prevalent and transmissible in Africa. A patient can be cured from these skin diseases if it is detected on time and treated at an early stage. However, it is difficult to identify these diseases in order to provide the right medications. This work presents an ensemble technique of five machine learning models to improve the accuracy level of classifying skin diseases. The result showed that the stacking method yield a high accuracy of 99.30% compared to the single classifiers and existing works.

In this work, we study how the performance of generative image models are impacted by the racial composition of their training datasets. By examining and controlling the racial distributions in various training datasets, we are able to observe the impacts of different training distributions on generated image quality and the racial distributions of the generated images. Our results show that the racial compositions of generated images successfully preserve that of the training data. However, we observe that truncation, a technique used to generate higher quality images, exacerbates racial imbalances in the data.

Schistosomiasis is a neglected disease affecting almost every region of the world, with its endemicity mainly felt in sub-Saharan Africa. A mathematical modelof Schistosomiasis integrating heterogeneous host transmission pathways is thus formulated and analyzed to investigate the impact of the disease on the human pop ulation. Mathematical analyses are presented, including establishing the existence and uniqueness of solutions, computation of the model equilibria, and the basic reproduction number (R0). Our numerical findings suggest that reducing the snail population will directly reduce Schistosomiasis transmission within the humanpopulation and thus lead to its eradication

Accurate and robust perception of the surrounding environment is a key component of all autonomous systems such as self-driving cars. Currently, widely used 3D object detectors employ complex handcrafted feature extractors and post-processors to produce semantic object interpretations from raw point clouds. In essence, the task of detecting 3D bounding boxes from point clouds can be reduced to a set-to-set transformation. The input is a set of points while the output is a set of bounding boxes. In this work, we streamline the 3D object detection pipeline by using a simple transformer architecture. We overcome the apparent challenge of quadratic memory and computation complexity of transformers by sampling the point cloud using a differentiable sampling network. We demonstrate the efficacy of our methods on the ubiquitous KITTI benchmark.

Music Information Retrieval (MIR) is the task of extracting high-level information, such as genre, artist, or instrumentation, from music. Genre classification is an important and rapidly evolving research area of MIR. To date, only a small amount of research has been done on the automatic genre classification of Nigerian songs. Accordingly, this study uses the k-nearest neighbors (k-NN), SVM, random forest (RF), and XGBoost classifiers [1] for music genre classification, due to the robust accuracy of its ensemble tree methods. These classifiers were applied to timbral and tempo characteristic features mined from 478 Nigerian songs from 5 music genres: apala (100), juju (120), fuji (99), Highlife (120), and waka (39). The objective was to assess the quality of music genre classification using the ORIN dataset, based only on the analysis of these features. SHapley Additive exPlanations (SHAP) [2] values with TreeExplainer (Tree SHAP) [3] were obtained to explain the model predictions and show feature importance in descending order. Usually, these orderings are different for the three options (weight, gain, and cover) used to measure feature importance by XGBoost classifier, but are consistent and accurate for Tree SHAP. Hence, the SHAP method avoids the inconsistency problem of current methods, therefore, increasing the power to detect true feature dependencies in a dataset and aiding the building of SHAP summary plots, which succinctly display the magnitude, prevalence, and direction of a feature’s effect [3]. The unique contribution of this study is threefold: (i) to build a new song dataset, ORIN, which will serve as an addition to the collection of publicly available MIR datasets; (ii) to build an automatic form of music genre classification for Nigerian songs that can support or replace the manual method; and (iii) to introduce the global mean Tree SHAP method to show feature importance and impact on the classification model’s output.

Explanatory Interactive Learning (XIL) collects user feedback on model explanations to implement a Human-in-the-Loop (HITL) based interactive learning scenario. Although XIL has been used to improve classification performance in multiple domains, the impact of different user feedback types on model performance and explanation accuracy is not well studied. To guide future XIL work we compare the effectiveness of two different user feedback types in image classification tasks: (1) instructing an algorithm to ignore certain spurious image features, and (2) instructing an algorithm to focus on certain valid image features. We show that identifying and annotating spurious image features that a model finds salient results in superior classification and explanation accuracy than user feedback that tells a model to focus on valid image features.

The new generation of radio telescopes, such as the MeerKAT, ASKAP (Australian Square Kilometre Array Pathfinder) and the future Square Kilometre Array (SKA), are expected to produce vast amounts of data and images in the petabyte region. Therefore, the amount of incoming data at a specific time will overwhelm any current traditional data analysis method deployed. Deep learning architectures have been applied in many fields, such as social network filtering and medical image analysis. They have produced results which are comparable to human expert performance. Hence, it is appealing to apply it to radio astronomy data. The images from these telescopes have a high density of radio sources, making it difficult to classify the sources in the image. Identifying and segmenting sources from radio images is a pre-processing step before sources are put into different classes. Thus, it is necessary to automatically segment the sources from the images before they can be classified. This work uses the Unet architecture to segment radio sources from radio images with 99.6 % accuracy. Thereafter, we use OpenCV tools to detect the sources and draw borders around them. PyBDSF and Unet were compared in the same environment ( computing power, images and dataset size), and it occurred that Unet is $35$ times faster than pyBDSF. This is the unique selling point of traditional vs deep learning approaches to radio images. Since MeerKAT is expected to produce a catalogue of about 70 million galaxies, this tool will speed up the runtime to produce this catalogue. The limitation of this work is that it uses pyBDSF as the ground truth image and cannot outperform it. For future improvements, hand-crafted mask images will be created as ground truth.

We share our methodology and findings from applying named entity recognition (NER) using machine learning, to identify behavioural patterns in transcribed, anonymized, privacy-preserving, and de-identified audio data from a customer contact centre data Nigeria based on the Fogg Behaviour Model (FBM). This work is part of a larger project that is focused on the practical application of AI to analyse and derive insight from large-scale data call centre data.The Fogg Behaviour Model (FBM) describes the interaction of three key elements (Motivation, Ability and a Prompt) and their interaction to produce behavioural change in relation to the adoption of positive reproductive health behaviour. This work is part of a larger project that is focused on the practical application of artificial intelligence to analyse and derive programmatic insight from large-scale customer contact centre audio data. The entity recognition model called Fogg Model Entity Recognition (FMER) was trained using spaCy, an open source software library for advanced natural language processing, on a total of 11510 words and scored an F1 of 98.5

In order to prepare for or respond to a humanitarian disaster, crisis, and emergency, public officials and media organizations often must disseminate large amounts of technical information in a short amount of time. As the result, misinformation can circulate within or outside the affected community, and such misinformation can be particularly deadly during disaster scenarios. Therefore, it becomes a challenge for public safety agencies and organizations to reduce or eliminate the spread of misinformation on social media. The modern era of Artificial Intelligence (AI) based fact-checking models relies on machine learning (ML) models to detect misinformation using sophisticated algorithms. However, most of these ML approaches are limited by the data used to train them and they are over-dependent on being accessed via smartphone app interface which may be insufficient in low income countries, where more than eighty percent of the population do not have smartphones. Hence in this paper, we propose an integrated fact-checking system that relies on a large network of independent and crowdsourced volunteer “checkers” who collect, verify and upload any fake messages into an app, which also has the functionalities to offer anyone the ability to verify any message they have received. The app can receive messages for verification in form of short message system (SMS), app, and chatbot. In order to address the challenge of high illiteracy level in low-income countries, this platform is also able to support basic featurephones via an SMS-based interface where “verifiers” can send any suspicious message to a dedicated phone number as SMS and receive instant call alert which confirms the veracity or otherwise of the message. The platform relies on the shared intelligence of the “crowd” to train the machine learning models for a higher degree of accuracy, while also offering an instant automated call service, which is available as pre-recorded messages in twenty local languages.

Building identification is an important task for urban planning, settlement tracking, and can also help to supplement the limited data in developing countries where there is inadequate and infrequent census data. Several Deep learning architectures such as Fully connected network (FCN), UNET and Deeplab can be used to perform building identification in such scenarios where census data is limited and have given promising results. However, most of these architectures have some drawbacks such as poor edge detection thus necessitating the use of very huge training datasets that in turn leads to the utilization of a lot of computation resources. Additionally, there is a challenge when it comes to adapting these trained models to other domains, i.e., a model trained in one region poorly performs on other regions. This research aims to conduct a comparative study of the different architectures used for building identification.

Existing federated learning uses a central server prone to communication and computational bottlenecks. Also, most existing federated learning algorithms do not cater for situations where the data distribution is time-varying such as in real-time traffic monitoring. To address these problems, this paper proposes a novel differentially private online mirror descent algorithm. To provide additional privacy to the loss gradients of the clients, local differential privacy is introduced.

Tuta absoluta is a major threat to tomato production, causinglosses ranging from 80% to 100% when not properly managed.Early detection of T. absoluta’s effects on tomato plants isimportant in controlling and preventing severe pest damageon tomatoes. In this study, we propose semantic and instancesegmentation models based on U-Net and Mask RCNN, deepConvolutional Neural Networks (CNN) to segment the effects ofT. absoluta on tomato leaf images at pixel level using field data.The results show that Mask RCNN achieved a mean AveragePrecision of 85.67%, while the U-Net model achieved anIntersection over Union of 78.60% and Dice coefficient of82.86%. Both models can precisely generate segmentationsindicating the exact spots/areas infested by T. absoluta intomato leaves. The model will help farmers and extension officersmake informed decisions to improve tomato productivityand rescue farmers from annual losses.

Glioma is one of the most deadly types of cancer diseases. Accurate assessment of pre-operative grading of glioma for patients with this disease can lead to better patient management. While Biopsy is the most commonly used diagnostic technique in routine clinical applications of glioma grading, it has several disadvantages such as it is invasive and prone to tissue trauma. Consequently, automated pre-operative glioma grade prediction techniques based on MR images are recently getting attention, so noninvasive. However, most of the recently developed automated techniques are based on the handcrafted image features extracted from the manually segmented tumor regions in MRI, which is tedious, time-consuming, and inaccurate. In this paper, we presented a novel automated glioma grade prediction method based on an ensemble of deep learning models. The proposed ensemble method is introduced in two main steps. In the first step, a novel deep learning architecture using the feature extraction layers of a pre-trained ImageNet model as a backend was proposed, and subsequently trained separately using 2D images reconstructed from the multi-modal MR images in the axial, coronal and sagittal planes, resulting in multiple base learners. In the second step, the outputs of the base learners were combined using various fusing strategies, including averaging, voting, and classical machine learning techniques. Experimental results on the BraTS benchmark dataset demonstrate that the proposed ensemble learning approach achieved superior performance compared to state-of-the-art MRI-based glioma grade prediction methods. We believe that the method proposed in this study can be used as a supporting tool for neurologists and radiologists in the the precise diagnosis of glioma.

AbstractArtificial intelligence has an increasingly large impact on everything from social media to healthcare. Artificial Intelligence is used to make credit card decisions, to conduct video surveillance in airports, and to perform military operations. These technologies have the potential to harm or help the people that they serve. This paper aims to identify some ethical challenges that affect data-driven policing in modern society and some solutions that have helped address some of these challenges. The police department plays a very huge role in keeping our communities safe, and using data to aid their decisions could cost us a lot if not properly looked at. ML applications have discriminated against individuals on the basis of race, sex, religion, socioeconomic status, and other categories. These could have severe consequences if some of these challenges are not examined.

End-to-End (E2E) approach to Automatic Speech Recognition (ASR) is a hot research agenda. It is interesting for less-resourced languages (LRL) since it avoids the use of pronunciation dictionary. However, E2E is data greedy, which makes the application of E2E to LRL questionable. However, using data from other languages in a multilingual (ML) setup is being applied to solve the problem of data scarcity. We have conducted ML E2E ASR experiments for four less-resourced Ethiopian languages using different language and acoustic modelling units. The results of our experiments show that relative Word Error Rate (WER) reductions (over the monolingual E2E systems) of up to 29.83% can be achieved by just using data of two related languages in E2E ASR system training. Moreover, we have also noticed that the use of data from less related languages also leads to E2E ASR performance improvement over the use of monolingual data.

The Cut-Test technique employs visual inspection of interior colouration, compartmentalization, and defects of beans for effective classification of cocoa beans. However, due to its subjective nature and natural variances in visual perception, it is intrinsically limited, resulting in disparity in the verdict, imprecision, discordance, and time-consuming and labor-intensive classification procedure. Although machine learning (ML) techniques have been proposed to fix these challenges with significant results, there is a need for improvement. In this paper, we propose a color and texture extraction technique for image representation as well as a generalized, less complex, and robust Neural Network model to help improve the performance of machine classification of Cut-Test cocoa beans. A total of 1400 beans were classified into 14 grades. Experimental results on the equal cocoa cut-test dataset, which is the standard publicly available cut-test dataset, show that the novel extraction method combined with the developed artificial neural network provides a more homogeneous classification rate for all the cocoa grades. The proposed model outperformed the Support Vector Machine, Decision Tree, Random Forest, and Nave Bayes on the same dataset. The proposed techniques in this work are robust on the cut-test dataset and can serve as an accurate computer-aided diagnostic tool for cocoa bean classification.

The goal of this project was to use deep learning to extract and isolate a mixed audio file. A specific sound of interest nearly often overlaps with other waves from other sources, and if those waves are at a similar or greater amplitude, it will hinder the listener's ability to perceive the sound properly. We can concentrate on the specific sound of interest with the help of audio separation and extraction using deep neural network called CNN. The raised CNN autoencoder model has two convolution layers for the encoding stage, one fully connected layer for the next step, two deconvolution layers for each class, and an array of completely connected layers for the decoding stage.

Imitation from observation is a paradigm that consists of training agents using visual observations of expert demonstrations without direct access to the actions.One of the most common procedures adopted to solve this problem is to train a reward function from the demonstrations, but this task still remains a significant challenge.We approach this problem with a method of agent behavior representation in a latent space using demonstration videos.Our approach exploits recent algorithms of contrastive learning of image and video and uses a bootstrapping method to progressively train a trajectory encoding function with respect to the variation of the agent policy. This function is then used to compute the rewards provided to a standard Reinforcement Learning (RL) algorithm.Our method uses only a limited number of videos produced by an expert and we do not have access to the expert policy function.Our experiments show promising results on a set of continuous control tasks and demonstrate that learning a behavior encoder from videos allows for building an efficient reward function for the agent.

Supervised machine learning (ML) algorithms are efficient at predicting the occurrence of diseases and they have become more popular as a result of the recent pandemic. A global re-emergence of measles has been reported and with the help of complete vaccination measles can be prevented. An average accuracy score of 0.90 confirms the predictive capacity of ML models. In terms of performance RFC, LDA and LR performed better than CART and KNN.

In this work we proposed a three-stages framework for real-time vehicles detection, tracking and speed estimation in the MLK Smart Corridor (A testbed for smart cities and connected/automated vehicles in downtown Chattanooga). YoloV4 is used to detect vehicles throughout successive frames. The Siamese network was trained with UA-Detrac and MLK Smart Corridor data. The proposed tracker was tested on MLK Smart Corridor data and scored 58.53% MOTA, ~2% percent more than Deep Sort, and the number of the identity switches have been reduced by 57%. Linear transformation is applied on the camera view to get a bird-eye view, from which we can estimate the corresponding meter-per-pixel (MPP). MPP along with the fps are used to get the speed directly. A vehicle with an On-Board Unit (OBU) is used to generate speed ground truth data. This generated data have been verified using a dashboard camera. The OBU readings' error rate was found to be ~0.89 mph. The speed estimator showed an average error rate of 2 mph on average traffic readings in the range of (0-35) mph.

The need to prevent blind planting and reduce farming time complexity is currentlymotivating precision agriculture to explore artificial intelligent techniques that wouldintroduce soil nutrients and location weather integration for crop yield prediction. Thiswork proposes the use of seven features to predict crop yield using a universal webapplication.

Aligning with the important theme of ``Diversity and Inclusion in Artificial Intelligence", we discuss the major linguistic, economic, and sociopolitical challenges facing development of Artificial Intelligence (AI) technologies for Africa and Africans. Focusing on speech and language processing, we discuss how the particulars of African languages and the context within which these are taught and used (or lack thereof) can be harnessed for technology development. Our main objective is to motivate and advocate for policies geared towards an Afrocentric approach to AI. With this in mind, we recommend what technologies to build and how to build, evaluate, and deploy them based on the needs of African communities.

Recently, deep learning based methods have been proposed to utilize the featureextraction power of deep neural networks to process bug reports. In this paper, we investi-gate Deep Ensemble learning for binary severity classification of bug reports. Due to thesignificance of feature representation for this task, we initially evaluate the effectivenessof word embedding models FastText and Word2Vec. We observe that the FastText modelachieves more generalised results; hence we adopt this model for the Deep Ensemble learningexperiments. We train and evaluate CNN, LightGBM, XGBoost, AdaBoost, and hybridmodels on seven data sets from Eclipse and Mozilla projects. Results analysis shows thatCNN LightGBM consistently outperforms CNN XGBoost by 9.89%, 5.89%, 10.1%, and8.16%, and CNN AdaBoost by 23.92%, 9.23%, 15.31%, and 11.1% in average accuracy,precision, recall, and f-measure, respectively. Similarly, the CNN LightGBM surpasses state-of-the-art approach by 10.69%, 13.6%, 0.16%, and 6.47% in average accuracy, precision,recall, and f-measure, respectively. Overall, the performance results demonstrates that ourchoice of the CNN’s depth with small hyperparameter tuning is a suitable approach. Italso shows that replacing the weak softmax classifier with a more powerful gradient boostclassifier enhances bug severity classification.

Adversarial learning represents a powerful technique for generating data statistics. Its successful implementation in quantum computational platforms is not straightforward due to limitations in connectivity, quantum operation fidelity, and limited access to the quantum processor for statistically relevant results. Constraining the number of quantum operations and providing a design with a low compilation cost, we propose a quantum generative adversarial network design that uses real Hilbert spaces as the framework for the generative model. We consider quantum generator and discriminator architectures based on a variational quantum circuit. For low-depth ans\"atze designs, we consider the real Hilbert space as the working space for the quantum adversarial game. This architecture improves state-of-the-art quantum generative adversarial performance while maintaining a shallow-depth quantum circuit and a reduced parameter set. We tested our design in a low resource regime, generating handwritten digits with the MNIST as the reference dataset. We could generate undetected data (digits) with just 15 epochs working in the real Hilbert space of 2, 3, and 4 qubits. Our design uses native quantum operations established in superconducting-based quantum processors and is compatible with ion-trapped-based architectures.

Building a usable radio monitoring automatic speech recognition (ASR) system is a challenging task for under-resourced languages and yet this is paramount in societies where radio is the main medium of public communication and discussions. The main challenge is the absence of transcribed radio speech datasets. In this paper, we create a Luganda radio dataset and build a COVID-19 ASR. We use the ASR to analyse public radio discussions for public health response. We openly release a radio speech corpus of 155 hours. To our knowledge, this is the first publicly available radio dataset in sub-Saharan Africa.

Airline disruption management traditionally seeks to address three problem dimensions: aircraft scheduling, crew scheduling, and passenger scheduling, in that order. However, current efforts have, at most, only addressed the first two problem dimensions concurrently and do not account for the propagative effects that uncertain scheduling outcomes in one dimension can have on another dimension. In addition, existing approaches for airline disruption management include human specialists who decide on necessary corrective actions for airline schedule disruptions on the day of operation. However, human specialists are limited in their ability to process copious amounts of information imperative for making robust decisions that simultaneously address all problem dimensions during disruption management. Therefore, there is a need to augment the decision-making capabilities of a human specialist with quantitative and qualitative tools that can rationalize complex interactions amongst all dimensions in airline disruption management, and provide objective insights to the specialists in the airline operations control center. To that effect, this paper provides a demonstration of an agnostic and systematic paradigm for enabling expeditious simultaneously-integrated recovery of all problem dimensions during airline disruption management, through an intelligent multi-agent system that employs principles from artificial intelligence and distributed ledger technology. Results indicate that our paradigm for simultaneously-integrated recovery executes in polynomial time and is effective when all the flights in the airline route network are disrupted.

This study aims to predict the level of anemia among pregnant women in the case of Ethiopia using homogeneous ensemble machine learning algorithms and deploy them on Heroku-based cloud computing for potential users. In this study, the data were gathered from the Ethiopian demographic, health survey (EDHS) collected three times at five-year intervals. The data were preprocessed to get quality data that are suitable for the machine learning algorithm to develop a model that predicts the levels of anemia among pregnant. The study was conducted following a design science approach. Random forest, cat boost, and extreme gradient boosting with class decomposition (one versus one and one versus rest) and without class decomposition were employed to build the predictive model. For constructing the proposed model, nine experiments were conducted with a total of 29104 instances with 23 features, and a training and testing dataset split ratio of 80/20. The overall accuracy of random forest, extreme gradient boosting, and cat boost without class decompositions are 91.34%, 94.26%, and 97.08.90%, respectively. The overall accuracy of random forest, extreme gradient boosting, and cat boost with one versus one are 94.4%, 95.21%, and 97.44%, respectively. The overall accuracy of random forest, extreme gradient boosting, and cat boost with one versus the rest are 94.4%, 94.54%, and 97.6%, respectively. Finally, the researcher decided to use cat boost algorithms with one versus the rest for further use in the development of artifacts, model deployment, risk factor analysis, and generating rules because it has registered better performance with 97.6% accuracy. We identified the most determinant risk factors using feature importance. Some of them are the duration of the current pregnancy, age in 5-year groups, source of drinking water, respondent's occupation, number of household members, wealth index, husband/partner's education level, and birth history.

Due to the limited availability for any researchers are interested in categorizing textdata using unseen labels. This became more feasible with the advent of Transformermodels. Many pre-trained models have been fine-tuned on entailment sentencepairs to perform dataless text classification with much success. However, otherresearchers have discovered that these large language models contain gender andracial biases that can negatively perpetuate negative stereotypes. While manyresearchers have explored the prevalence of gender bias in pre-trained word andsentence embeddings, there hasn’t been much research done in measuring andmitigating gender bias in zero-shot text classification. In this research, I propose amethod for evaluating gender bias in zero-shot text classification models and applythis technique on BART.

Algorithmic pricing on online e-commerce platforms raises the concern of tacit collusion, where reinforcement learning algorithms learn to set collusive prices in a decentralized manner and through nothing more than profit feedback. We demonstrate that reinforcement learning (RL) can also be used by platforms to learn buy box rules that are effective in preventing collusion by RL sellers and to do so without reducing consumer choice. For this, we adopt the methodology of Stackelberg POMDPs, and demonstrate success in learning robust rules that continue to provide high consumer welfare together with sellers employing different behavior models or having out-of-distribution costs for goods.

Numerous works in image-to-image translation have leveraged Generative Adversarial Networks (GANs) on unpaired datasets. As far as video translation is concerned, current GAN-based approaches do not entirely leverage space-time knowledge in videos. This research examines the idea of using GANs for the utilization of spatial-temporal information in a video by extending the unpaired video-to-video translations to enhance spatial-temporal awareness by adding feature preserving loss and temporal aware discriminator to generate more temporal consistent videos. Extensive qualitative and quantitative assessments demonstrate the notable success of the proposed system against existing methods. This paper illustrates that adding feature preserving constraints and temporal aware discriminator does improve temporal coherency of generated output video.

Research regarding the problem of defective insulator recognition on power distribution networks retains an open interest, due to the significant role insulators play to maintain quality service delivery. Most existing methods detect insulators by rectangular bounding box but do not perform segmentation down to instance pixel-level. In this paper, we propose an automated end-to-end framework enabled by attention mechanism to enhance recognition of defective insulators. Using natural industry dataset of images acquired by unmanned aerial vehicle (UAV), pixel-level recognition is formulated into two computer vision tasks; object detection and instance segmentation. We increase the capabilities of our chosen model by leveraging a lightweight but effective three-branch attention structure integrated into the backbone network as an add-on module. Specifically, we exploit cross-dimensional interactions to build an efficient computation of attention weights across channels of the backbone network to achieve gains in detection performance for defective insulators up to about +2.0 points compared to our base model, at negligible overhead cost. Moreover, we implement a training scheme to improve segmentation performance while demonstrating segmentation superiority over traditional segmentation approaches.

dentifying differences between genomic data sets can help us study diseases better and find their treatments. In this project, we trained autoencoders on HiC-maps of human cells and tested them on colon cells. HiC maps are genomic wide contact maps that tell us the relationship each entity has with its environment. We experimented with a simplified and a regularized autoencoder which produced a result of 0.022 and 0.021 training losses respectively. We then integrated Concept saliency map to highlight pixels that were relevant for a prediction by our model. Our model was able to highlight key areas of differences between a treated and an untreated cancerous human colon cell.

The epidemic Vulnerable-Latent-Contagious-Recovery-Inoculation (VLCRV-I) was proposed. Thereafter, an equivalent multi agent model was developed in order to cater for malware spread in computer networks. Then, various LSTM types was used for prediction and metrics such as Mean Squared Error, Root Mean Squared Error, Mean Absolute Error, Mean Absolute Percentage Error were used to evaluate model performance. The best prediction of vulnerable computers were obtained using Stacked LSTM of 512 Layers and the Relu Activation Function.

Ensemble learning techniques have achieved state-of-the-art performance in diverse machine learning applications by combining the predictions from two or more base models. This paper presents a concise overview of ensemble learning, covering the three main ensemble methods: bagging, boosting, and stacking, their early development to the recent state-of-the-art algorithms. The study focuses on the widely used ensemble algorithms, including random forest, adaptive boosting (AdaBoost), gradient boosting, extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), and categorical boosting (CatBoost). An attempt is made to concisely cover their mathematical and algorithmic representations, which is lacking in the existing literature and would be beneficial to machine learning researchers and practitioners.

Generating realistic simulated data for evaluating algorithms in healthcare remains a challenge as expert-based models overestimate performance on ML tasks, while data-driven models like GANs do not allow for ablation studies. To address this, we propose an approach that learns the properties of real time series, then augments simulated data with them. On glucose forecasting, we show that our method brings performance closer to that of real data compared to current simulation practices.

Despite the various intervention by the government and private bodies to help the agriculture sector, farmers in Nigeria still complain about poor productivity. Utilizing technology, research, and development to the fullest extent is necessary for sustainable agricultural practices. This study aims to demonstrate the advancement and accuracy of machine learning algorithms in predicting the best crop to be grown based on field data. The outcome of the various algorithms used resulted in high prediction, driving home the suitability of machine learning algorithms in improving agricultural practices and methods. Adopting precision farming will have an exponentially positive effect on crop growth and crop productivity, as farmers will make informed decisions thereby increasing the sustainability of the agricultural sector and ultimately strengthening the economy.

We have seen significant growth in the area of building Natural Language Processing (NLP) tools for African languages. However, the evaluation of gender bias in the machine translation systems for African languages is not yet thoroughly investigated. This is due to the unavailability of explicit text data available for addressing the issue of gender bias in machine translation. In this paper, we use transfer learning techniques based on a pre-trained Marian MT model for building machine translation models for English-Luganda and Luganda-English. Our work attempts to evaluate and quantify the gender bias within a Luganda-English machine translation system using Word Embeddings Fairness Evaluation Framework (WEFE). Luganda is one of the languages with gender-neutral pronouns in the world, therefore we use a small set of trusted gendered examples as the test set to evaluate gender bias by biasing word embeddings. This approach allows us to focus on Luganda-English translations with gender-specific pronouns, and the results of the gender bias evaluation are confirmed by human evaluation. To compare and contrast the results of the word embeddings evaluation metric, we used a modified version of the existing Translation Gender Bias Index (TGBI) based on the grammatical consideration for Luganda.

In the present, the growth of digitization and worldwide communications make2 OCR systems of exotic languages a very important task. In this paper, we develop an OCR model for one of these exotic languages with a unique script, Amharic. Motivated with the recent success of the Attention mechanism in Neural Machine Translation (NMT), we extend the attention mechanism for Amharic text-image recognition. The proposed model consists of CNNs and attention embedded encoder-decoder networks that are integrated following the configuration of the seq2seq framework. Unlike the existing OCR model that minimizes the CTC objective function, the new model minimizes the categorical cross-entropy loss. The performance of the proposed attention-based model is evaluated against the test dataset from the ADOCR database which consists of both printed and synthetically generated Amharic text-line images, and achieved a promising result with a 13 Character Error Rate (CER) of 1.54% and 1.17% respectively.

In recent times, many researchers have explored the prediction of infectious disease outbreaks using mathematical modeling, artificial intelligence, agent-based simulation models among other technologies. However, there is still a need to improve on the prediction accuracy of the epidemiological surveillance systems–since the emergence of outbreaks is due to the multi-level interactions of humans, pathogens, and environments. Cogent socio-ecological research efforts suggest that the phenomenon of infectious disease outbreaks is best tackled from the perspective of Complex Adaptive systems (CAS). In this study, we provided a scoping review of various approaches adopted in the literature for epidemiological surveillance systems– with the goal of creating a new pathway for a more robust surveillance model using a deep learning approach enhanced with an equilibrium state bifurcation technique for early, and a more accurate detection of infectious disease outbreaks in epidemiological surveillance systems.

Automatic speaker recognition is the task of automatically determining the identity of a speaker from a recording of their voice sample. One of the most important steps of speaker recognition that significantly influences the speaker recognition performance is known as feature extraction. Acoustic features of speech have been researched by many researchers around the world, however, there is limited research conducted on African indigenous languages. This paper presents the effects of acoustic features of speech towards the performance of speaker recognition systems focusing on South African low-resourced languages. This study investigates three acoustic features of speech namely, Time-domain, Frequency-domain and Cepstral-domain features extracted from the National Centre for Human Language Technology (NCHLT) Sepedi speech data. The results show that the performance is poor for time-domain features and good for frequency-domain features and even better for cepstral-domain features. However, the combination of these three features resulted in a higher accuracy.

ABSTRACTSpeech has been widely known as the primary mode of communication among individuals and computers. The existence of technology brought about human computer interface to allow human computer interaction. While Speech Emotion Recognition Systems have been developed at a rapid pace over the last few years a lot of challenges has also been encountered during the course of this development such as inability to detect emotions that causes depression and mood swings which can be used by therapists to monitor the moods of their patients. To enhance the power of Speech Emotion Recognition models, it is required to design a model that recognizes the different emotions that leads to depression to enhance doctor-patient relationship. In this paper, we used the Knowledge Discovery Database (KDD) methodology and the features extracted were Zero Crossing Rate (ZCR), Mel-Frequency Cepstral Coefficients (MFCC) and Root Mean Square (RMS). The model was built using Tensorflow CONV1D with relu activation function and multiple sequential layers. An epoch size of 50 and a batch size of 64 were used. The result shows that using confusion matrix as the performance metrics yielded an accuracy of 96%.

Cassava is a staple crop that is important for food safety in parts of Africa. A key challenge in growing the crop is that it is highly sensitive to diseases. Today, experts primarily diagnose these diseases by moving to different parts of the country while visually assessing the state of health of the crops, which is a cumbersome and erratic process. Nevertheless, state-of-the-art deep transfer learning models that can aid the automated diagnosis of these diseases exist. However, these models cannot be deployed on mobile devices because of the limited memory and computational capacity of these devices and there is not enough network coverage to service them from the cloud. To address this issue, we present knowledge distillation as a technique that can be used to build accurate plant disease classification models that are compatible with the capabilities of mobile devices. We train new Mobile-PDC Plant Decisive Classification models that have the same classification accuracy as state-of-the-art PDC models, but are much smaller in size and fit on mobile devices. Our Mobile-PDC models have the MobileNet structure, which makes them compatible with multiple mobile devices. Our experiments demonstrate that we can compress 91.2% of the original state-of-the-art PDC models without losing accuracy.

Current Convolutional Neural Networks (CNNs) does not explicitly capture and select diverse image features. Rather follow an indirect approach of increasing the networks' depth or width, which significantly increase the computational cost of the models. Inspired by biological visual system, this paper proposes a Multi-Kernel Attention Convolutional Network (MKA-Net ), which enables any feed-forward CNNs to explicitly capture and select diverse informative features to efficiently boost CNNs' performance. MKA-Net infers attention from the intermediate feature map by first using multiple sizes of kernels to capture diverse features then exploit neighboring feature-map relationship to adaptively select the most informative features. MKA-Net incurs negligible computational overhead and is designed to be easily integrated with any CNN architecture. We extensively evaluated the proposed MKA-Net module on benchmark datasets, including CIFAR100, SVHN, and ImageNet, with various CNN architectures. The experimental results show our approach provides a significant performance improvement with very minimal computational overhead.

Automated detection and localization of surface defects is critical for the timely maintenance and repair of planar materials-based industrial products from the automobile to the aerospace industry. This paper conduct a detailed and systematic comparative analysis of various anchor-based and anchor-free DL-based algorithms. The experimental results are further analyzed using the mean Average Precision (mAP) value and the impact to which augmentation strategies generalizes the model performance. The comparative analysis study presented in this paper helps to gain insight into the strengths and limitations of the popular DL-based frameworks under practical constraints with their real-time deployment feasibility.

Self-supervised contrastive learning is among the recent representation learning methods that have shown performance gains in several downstream tasks including semantic segmentation. This paper evaluates strong data augmentation, one of the most important components for self-supervised contrastive learning's improved performance. Strong data augmentation involves applying the composition of multiple augmentation techniques on images. Surprisingly, we find that the existing data augmentations do not always improve performance for semantic segmentation for medical images. We experiment with other augmentations that provide improved performance.

This paper presents“DeepWare,” a ransomware detection model inspired by deep learning and hardware performance counter (HPC). By imaging the HPC values and restructuring the conventional CNN model, DeepWare can address HPC’s nondeterminism issue by extracting the event-specific and event-wise behavioral features, which allows it to distinguish the ransomware activity from the benign one effectively. The experiment results across ransomware families show that the proposed DeepWare is effective at detecting different classes of ransomware with a 98.6% recall score, which is 84.41%, 60.93%, and 21% improvement over RATAFIA, OC-SVM, and EGB models, respectively.

This paper examines the viability of deep learning models for Afaan Oromo hate speech recognition. Toward this, we collect and annotate the first and most enormous Afaan Oromo language social media datasets. Variations of profound deep learning models such as CNN, BiLSTMs, LSTM, GRU, and CNN-LSTM are examined to evaluate their viability in identifying Afaan Oromo Hate speeches. The examination result uncovers that the model dependent on CNN and Bi-LSTM outperforms every one of the models on the test dataset with an average F1-score of 87%. Overall, considering the nature of the Afaan Oromo language and the prevalence of hate speech, we believe this study’s finding is promising for future works.

Sentiment lexicon plays a vital role in lexicon-based sentiment analysis. The lexicon-based method is often preferred because it leads to more explainable answers in comparison with many machine learning-based methods. However, lexicons that include only unigrams do not capture contextual information. To this end, we automatically generate domain-specific lexicons and manually develop contextual shifters. We show that for sentiment classification tasks in the economics and finance domain, the symbolic approach achieves competitive performance as the deep neural network. In addition, the symbolic approach provides understandable explanations. We will release the lexicons, shifter patterns and models to motivate future research in this direction.

This work proposes to solve the problem of few-shot biometric authentication by computing the Mahalanobis distance between testing embeddings and a multivariate Gaussian distribution of training embeddings obtained using pre-trained CNNs. Experimental results show that models pre-trained on the ImageNet dataset significantly outperform models pre-trained on human faces. With a VGG16 model, we obtain a FRR of 1.18% for a FAR of 1.25% on a dataset of 20 cattle identities.