We are aware that AI can be used for a variety of applications. The phrase AI for social good has become more prominent over the past few years. But what is meant by AI for social good? AI for social good is a newer research field that has an emphasis on using AI to address social, environmental, and health challenges. This research is multiple disciplinary which can help provide a variety of techniques and perspectives to address these challenges. On the surface this sounds promising. However, will some communities more than others have the opportunity to benefit from this research and in what ways? For instance, can AI for social good decrease trolls in social media spaces of members of the African diaspora or determine trends or patterns regarding Black maternal mortality? The focus of this talk will explore and try to determine what does AI for social good mean specifically to members of the African diaspora.

Machine learning techniques for analyzing human language have traditionally focused on English, other European languages and widely-spoken Asian languages. While African languages have been getting more attention recently, the same forces that led to them being ignored in the first place are being replicated internally within Africa. Landmark studies in African Natural Language Processing (NLP) focus on widely spoken languages like Swahili, Yoruba, Amharic, Zulu, etc. Languages with fewer speakers, such as all Ghanaian languages, are relegated to the "Future Work" section and are rarely treated as a priority by researchers, research funding agencies or press. I will discuss NLPGhana, an initiative I co-founded to ensure that Ghanaian languages are treated as a priority in African NLP research. I will discuss some of our technical achievements - such as Khaya, the world's first and only Ghanaian language neural machine translation app that is already being used widely by the Ghanaian public. I will discuss challenges faced in trying to make this research ecosystem self-sustaining, given the aforementioned lack of support for languages with fewer speakers.

API management tools to make secure, traceable software practically accessible to everyone

Actionable intelligence for smallholder farmers in Uganda

We enable young Africans to code using smartphones

The legal domain is attracting considerable attention in natural language processing (NLP) due to the number of legal documents generated (contracts, business deals, etc.) throughout professional activities and the logical business processing required on that documents. Treat legal documents is particularly cumbersome due to the context-specific knowledge and its extensive length. BigBird has achieved significant performance both on the computational side and on learning representation in the long-range arena. Few researchers have investigated the ability of long-range Transformer models to tackle the knowledge representation problem in the legal domain. We present in this work an adaptation of the long-range Transformer-based model BigBird on legal domain complemented with a use case in legal case retrieval. We continued the training of BigBird with the self-supervised learning task masked language modeling on legal corpora. Without fine-tuning, we tested the pre-trained models on legal case retrieval. We showed that adapting BigBird on legal corpora improves the knowledge representation of documents and outperforms by 5 in accuracy score the vanilla BigBird on the same task.

We introduce a way to get a chatbot to improve using a unique type of reinforcement learning. We get the chatbot itself to evaluate its responses and indicate alternate responses that would be better in quality.Here both the actor and the critic are the same system. We then teacher force the better response against the utterance that was parsed to the chatbot. Our experiments show that this may be a good way to optimize a chatbots ”policy”.

In recent years, there has been a resurgence in research on empirical methods for machine translation. Most of this research has been focused on high-resource, European languages. Despite the fact that around 30% of all languages spoken worldwide are African, the latter have been heavily under investigated and this, partly due to the lack of public parallel corpora online. Furthermore, despite their large number (more than 2,000) and the similarities between them, there is currently no publicly available study on how to use this multilingualism (and associated similarities) to improve machine translation systems performance on African languages. So as to address these issues, we propose a new dataset (from a source that allows us to use and release) for African languages that provides parallel data for vernaculars not present in commonly used dataset like JW300. To exploit multilingualism, we first use a historical approach based on migrations of population to identify similar vernaculars. We also propose a new metric to automatically evaluate similarities between languages. This new metric does not require word level parallelism like traditional methods but only paragraph level parallelism. We then show that performing Masked Language Modelling and Translation Language Modeling in addition to multi-task learning on a cluster of similar languages leads to a strong boost of performance in translating individual pairs inside this cluster. In particular, we record an improvement of 29 BLEU on the pair Bafia-Ewondo using our approaches compared to previous work methods that did not exploit multilingualism in any way. Finally, we release the dataset and code of this work to ensure reproducibility and accelerate research in this domain.

In this paper, we introduce SautiDB-Naija, a speech corpus of non-native speakers of English intended for research in accent translation, voice conversion, pronunciation classification, and accent classification. This initial release of our corpus includes over 900 recordings of non-native speakers of English whose first language (L1) is amongst the most common in Nigeria, namely Yoruba, Igbo, Edo, Efik-Ibibio, and Igala. To the best of our knowledge, this would be the first documented effort to curate a corpus of Nigerian accents for machine learning research to date. We demonstrate that neural networks are capable of learning linguistic features that distinguish between different accent classes by training a discriminative classifier on our corpus. This demonstrates the potential of SautiDB-Naija as a valuable resource for future computational linguistic research.

Federated Learning (FL) has recently emerged as a privacy-preserving method to train a deep learning model in a decentralized fashion while keeping the training data on edge devices. In addition to privacy, FL eliminates the need to send and store a lot of data on a central server, often in the cloud. Many researchers suggest various FL algorithms, each dealing with a particular weakness of previous algorithms. Unfortunately, it is not easy to compare these algorithms because each researcher (1) has different implementations, (2) tests them in different environments, and (2) uses different evaluation metrics. In this work, we implement two of the most popular FL algorithms, Federated Average (FedAvg) and Federated Proximal (FedProx), test them in the same environment, and use the same evaluation metrics on an image classification task. We constructed four experiments, each with a different percentage of stragglers. We trained ten different models for image classification for each experiment. We observed that FedAvg achieves high accuracy while taking less time to train a model than FedProx in settings with few to no stragglers. However, with a high percentage of stragglers (up to 90%), our results show that FedProx outperforms FedAvg by achieving high accuracy on average. We also noticed that FedAvg is highly unstable in environments with a high percentage of stragglers compared to FedProx. Lastly, we observed that FedProx is robust to statistical and system heterogeneity, while FedAvg is less robust regarding system heterogeneity in environments with a high percentage of stragglers.

The Linear State-Space Layer (LSSL) is a model family that combines the strengths of sequential modeling paradigms such as recurrence, convolution, and differential equations. For example, they generalize convolutions to continuous-time, explain common RNN heuristics, and share features of NDEs such as time-scale adaptation. Although naive LSSLs struggle with modeling long dependencies, we introduce a class of LSSL (SLLSSL), which overcomes these limitations by utilizing a trainable set of structured matrices that endow it with long range memory.

We present ongoing research on using data cartography techniques to better understand the training dynamics of neural sequence-to-sequence models for Machine Translation (MT). This is particularly needed to inform the design of effective training algorithms for MT in low-resource settings. Current results show that the difficulty of training samples depends on the current training phase and raises questions for future work, i.e. what are the properties of these samples and how can we modify the training process at each phase.

Afrocentric fashion images suitable for machine learning tasks are not well represented in open datasets. In this work we present AFRIGAN, a generative adversarial model for contemporary african fashion images. AFRIGAN can be leveraged as a tool for realistic image data synthesis, design iteration and experimentation for contemporary African fashion styles. This model is openly available

We present AFRIFASHION40000, an openly available dataset of African fashion images generated using Generative Adversarial Networks (GANs). This work explores the practical application of Artificial Intelligence to contemporary African art for new designs, image data synthesis, and representation in computer vision tasks. This is a sequel to AFRIFASHION1600, a small dataset of contemporary African fashion images. AFRIFASHION40000 contains 40000 images with a dimension of 128 X 128 for each image in 8 classes of items. An interactive interface and download link are available

The ability to notice mispronunciations is a key skill for second language learners. Unfortunately, it usually difficult for learners to acquire consistent feedback concerning their current level of speaking and listening skill. This issue is exacerbated by the fact that standard systems usually accept one correct pronunciation, but humans can understand a larger range of pronunciations which include mispronunciations. Current ASR systems are able to recognize speech almost perfectly, but often don't work when the speaker is not native to that language, i.e. is not spot on with their pronunciation, and more importantly, they don't provide feedback to the user on how to move from the mispronunciation to a more understandable pronunciation, when detected. We propose an approach to detect mispronunciations using a Siamese Network that is trained to recognize not a single correct pronunciation but instead a range of pronunciations that are user defined. The user can control the range of tolerance within which a word is understandable and is based on phoneme pronunciations. As the user interacts, they can adjust the range of tolerance based on their current need.

With current medical interest in the minimally invasive surgery paradigm, there has been palpable interest in the nascent field of machine learning in robot-assisted surgery. Consequently, several attempts have been made to train neural networks that extract intelligible data from robotically controlled surgical procedures.

These attempts have focused on single data points (e.g. action recognition or surgical skills assessment), and have barely reached model training thresholds adequate enough to be deemed useful in the high-stakes surgical domain.

In this paper, we propose a neural network training regime that accounts for both surgical action recognition and surgeon skills assessment while also training above prior validation accuracy benchmarks. More specifically, we use an attention mechanism that mimics the visual perception attention mechanism humans use to solve domain specific tasks.

To incorporate an attention mechanism in the action recognition and skills assessment processes, our attention implementation simultaneously recognizes three information benchmarks: the visual information in each frame, knowledge of the ongoing task(s), and the spatial attention in previous frames. Our implementation resulted in a 20 percentage-point-average increase in top-1 validation accuracy of all surgical action recognition and skills assessment tasks.

In this paper, we have studied on real-time speech-to-speech translation model for Amharic and Afaan Oromo languages. The model studied has three basic components such speech recognition, machine translation, and speech synthesis. For the speech recognition section, we have used HMM, a Hybrid approach for MT, and a concatenative synthesizer for TTS translation. HTK for speech recognition, IRSTLM for language modeling, GIZA++ for word-level alignment, MOSES as decoder, and Festvox tool for speech synthesis are used as a toolkit. In our evaluation we have found the ASR mode with an accuracy of 89.21%, the MT module shows 90% accuracy for Amharic to Afaan Oromo translation and 88.4% Accuracy for Afaan Oromo to Amharic translation and, TTS synthesizer scores 3 out of 5 on average after it is evaluated by three individuals. Keyword: machine translation, speech recognition, speech synthesis, language model.

In recent years there have been great interests in addressing the low resourcefulness of African languages and providing baseline models for different Natural Language Processing tasks. Several initiatives on the continent use the Bible as a data source to provide proof of concept for some NLP tasks. In this work, we present the Lingala Speech Translation (LiSTra) dataset, release a full pipeline for the construction of such dataset in other languages, and report baselines using both the traditional cascade approach (Automatic Speech Recognition -> Machine Translation) and a revolutionary transformer-based End-2-End architecture with customized interactive attention that allows information sharing between the recognition decoder and the translation decoder.

When considering off-policy reinforcement learning methods for treatment policies in healthcare data, it is generally the case that the patient population is diverse and has different chronic conditions that we would like to take into account when identifying optimal treatment policies. In this work, we use multi-group Gaussian process regression models in a fitted Q-iteration framework to allow us to model these different patient subgroups and adapt the optimal policies to each subgroup while estimating these function across the entire patient population. We apply our multi-group reinforcement learning (MGRL) framework to the problem of optimal treatment policies for women in childbirth with pre-existing conditions and across ethnicities to show the performance against other state-of-the-art methods

Agricultural commodity price prediction helps the government, investors, and farmers to make informed decisions. Realizing the benefit, several researchers proposed different prediction models that use different features. However, most prediction models are affected by factors, such as data type (e.g., linear and nonlinear), seasonality of commodity items, weather conditions, commodity volatility features, and country economic factors. To solve this problem, we propose a model that combines time-series features and technical indicators to forecast the commodity price. The prediction model is created with four machine-learning algorithms: artificial neural network (ANN), support vector machine (SVM), random forest, and MLP regressor. The prediction model is built using four-machine learning algorithms namely, artificial neural network (ANN), support vector machine (SVM), random forest, and MLP regressor. To see the impact of combined features, we conducted two experiments using coffee and sesame datasets. The performance of the prediction models is assessed using the root mean square error (RMSE) and mean average error (MAE). The results show that the proposed model perform better than the baseline approach by an average by an average of 4.3753, 4.4216, 2.7494, and 6.658 while using Artificial Neural Networks, Support Vector Machines, MLP regressor, and Random Forest, respectively. To see which of the features contributed to the improvement of agricultural commodity price prediction, we computed feature importance using ReliefFAttributeEval. The result shows that: EMA, DEMA, SMA, True High, True Low, RSI, Trend, ADX, Seasonality (volatility), and CMO were founded in the top 10 in their predictive ability with the respective order.

The State Capture Commission of Inquiry has piqued the interest of a large number of viewers over the past few years. This research aims to help ease the burden of a user sifting through hundred of videos for a topic of interest through the proposal of a topic extraction based recommender system. With a large textual dataset at disposal, the research investigates the use of topic extraction through the combination of approaches such as Non-Negative Matrix Factorization and Latent Dirichlet Allocation for the construction of an algorithm that will be used for keyword based topic extraction. The result of this will allow the most relevant State Capture video related to a topic being presented to the user.

This paper presents a study on fine-tuning the pre-trained language models such as T5 to generate analytical textual summaries, describing the classification performance of machine learning models. The generation is based on the evaluation metrics achieved on a given classification problem. Evaluation of the generated metrics' narrations, indicates that exploring pre-trained models for data-to-text generation leads to better generalisation performance and can produce high-quality analytical summaries.

A major problem for High-Risk AI ( those used in health care, criminal justice, lending, housing, education, etc. ) is the need for regulatory oversight of model activity and end consumer user knowledge of when AI is being used. SquawkSierraTM proposes a central repository of ML model usage dynamically built by self referencing transponder models.

Facial expressions are an important means of non-verbal communication, sometimes serving as a complement to sending messages in human interaction. The objective of the present work is to carry out a comparison between three models of classification of emotions through the face, developed by different convolutional neural network architectures (CNN).

Designing robots that can learn multiple tasks is a difficult problem because learning from scratch for each task is prohibitively slow, leveraging experience from previous tasks might be difficult, and there might not exist a suitable reward function for RL. Thus some desirable properties are the ability to learn from demonstrations through imitation learning, which requires no reward function and can lead to exponential decreases in sample complexity, and to leverage past experience through meta-learning, which improves few-shot learning. Such robots would also benefit from leveraging hierarchy to re-use simple learned skills for more complex tasks, but learning hierarchical policies has historically proven difficult. In this work, we move closer towards satisfying these properties by formulating a hierarchical imitation learning problem which we tackle through meta-learning to leverage experience from multiple tasks. We present results on a linear point mass environment as well as a challenging simulated kitchen environment with a 7-DoF robotic arm.

Persons with disabilities in Ethiopia are not exceptions in experiencing the problems associated with the non-inclusiveness of AI technological devices. Some service providers in the country are using AI technologies. However, some of such AI devices are not disability friendly to access or to be accessed by persons with disabilities. As of today, it is hardly possible to find the tasks of human beings that are beyond the reach of science and technology. Nevertheless, it is not unusual to find artificial intelligent technological devices that are non-inclusive as to some groups of people like persons with disabilities. These artificial intelligent technological devises and systems, while they need to be more supportive and too much promising for people with disabilities, in reality, are sometimes found discriminatory or incompatible to be accessed or to access persons with disabilities. These problems are mainly because of some errors, by negligence – if not intentionally, during data setting, preparation, learning, model deployment, and implementation of such devises and systems. The non-inclusiveness of some of such artificial intelligent devices and systems is evidenced in different scenarios like HRM system for recruitment, face recognition based attendance system, supermarket intelligent CCTV camera, and many more aspects. This study, by assessing some non-inclusive artificial intelligent devices with a particular emphasize in Ethiopia, scrutinizes the operation of such devices as to the rights of persons with disabilities. Doing so, the study reveals the discriminatory nature of some of such devices and the corollary violations of human rights of persons with disabilities. And, it provides the ways forward to be considered.

The necessity to use non-invasive and fast methods for the diagnosis of pulmonary diseases is more pressing than ever with the emergence of the SARS-CoV-2. Given their documented performances on image classification tasks, deep learning algorithms constitute promising approaches that can rely on abundant data to achieve the detection of pulmonary diseases from chest X-rays. However, due to privacy issues and some patients refusing to disclose their chest X-rays, data can be scarce and hence prevent deep learning algorithms from achieving their optimal capacities. In this paper, we use a dataset containing a total of 270 training samples and 36 training samples with 4 classes (normal, bacterial pneumonia, viral pneumonia and COVID-19). Using the efficientnet-b5 architecture as the backbone, we investigate, by observing the model performances, if transfer learning can be used to overcome data scarcity when training a deep leaning model for a pulmonary disease detection task.

According to UNICEF, malaria kills a child every 2 minutes, especially in marginalised communities. Malaria mortality can be drastically reduced by ensuring prompt access to diagnosis and treatment. However, using the microscope, the recommended diagnosis tool, is expensive, expert dependent, time-consuming for a single diagnosis and becomes impractical in areas with a high disease burden. Traditional data labeling processes use tools such as Labelmg to manually annotate every object and this makes it an expensive process and time consuming thus leading to limited dataset especially in a marginalized community. In this research, we present findings on using deep learning techniques to facilitate a fast and effective creation of ground truth datasets to be used in developing relevant malaria diagnosis tools, drawing on data from Tanzania. Our results demonstrate that it took one third less time with high efficiency to annotate the dataset compared to traditional methods. This annotation technique provides the assurance of the availability of high-quality labeled malaria datasets that can be used to develop machine learning based malaria diagnosis tools.

Non-data mining techniques have been used to analyze malaria data and public health strategies have been implemented based on these analyses to limit the impact of malaria. However, malaria is still prevalent in our communities thus prompting further analyses of malaria data. This study adopts a descriptive data mining approach: association rule mining, to analyze malaria data and investigate relations that can explain the prevalence of malaria. The apriori algorithm was used, after converting the data from numeric to categorical data, and multidimensional association rule measure carried out on generated rules. The minimum support threshold was set at 0.3 and 0.02 for global and local investigations respectively, and minimum confidence set at 0.8 and 0.02 likewise. The results from this study showed a strong association between malaria diagnosis and malaria deaths, among patients aged 5 and above. In addition, synthetic data was generated from the real data using classification and regression trees, and the same investigations carried out. The results from the synthetic data showed a similar trend as that of the real data. Although children below the age of five and pregnant women have been the focus of public health action in my community, this study suggests that attention should also be paid to those above 5 years.

Alzheimer's disease (AD) is known to be gradual in its progression of irreversible neuronal damage and eventual death. Patterns of atrophy are strongly correlated with AD pathology, specifically, morphological changes in brain shape, which have been identified to occur up to ten years before clinical diagnoses. Structural neuroimaging modalities (e.g., MRI) make it possible to analyze brain shape using intermediate representations of 3D shape such as voxels and point clouds but typically suffer from high computational complexity and an absence of smoothness in 3D shape. We propose geometric deep learning models for analyzing AD pathology using graph neural networks composed of fast and efficient spiral mesh convolutional layers, which are trained on surface mesh representations of neuroanatomical structures. Our discriminative spiral network outperforms alternative methods and shape representations for AD classification. Our proposed generative mesh networks, conditioned on AD diagnosis, demonstrate volume and surface area reductions in subcortical regions affected by AD neurodegeneration as well.

Exposure to prosocial models is commonly used to foster prosocial behavior in various domains of society. Although their translational value has been challenged in several cases, rodents, in particular Long Evans rats, are invaluable when modeling the context of human behavior in alternate rearing-environments. Rodent models are chosen for convenience, i.e., availability rather than neurotechnologically enabled neurosurgical interventions involving the human brain. Apart from a general discussion on translational success or failure, the complex bidirectional process involving multidisciplinary research integration often requires new data science methods. This research innovates by intersecting the disciplines of neuroscience and artificial intelligence (AI). But the novel union of AI, data science, and neuro-behavioral experiments comes with benefits and the potential for harm. We conclude by reporting the appropriateness of the enhanced K4-RANN as an effective ML algorithm for estimating the nonlinear regression weights of neural activity in trait-bred Long Evans rats subjected to alternate rearing environments. However, the study finds this conclusion is most true under the principle of universal approximation for algorithmic parameter settings.

In this paper, we introduce SautiDB-Naija, a speech corpus of non-native speakers of English intended for research in accent translation, voice conversion, pronunciation classification, and accent classification. This initial release of our corpus includes over 900 recordings of non-native speakers of English whose first language (L1) is amongst the most common in Nigeria, namely Yoruba, Igbo, Edo, Efik-Ibibio, and Igala. To the best of our knowledge, this would be the first documented effort to curate a corpus of Nigerian accents for machine learning research to date. We demonstrate that neural networks are capable of learning linguistic features that distinguish between different accent classes by training a discriminative classifier on our corpus. This demonstrates the potential of SautiDB-Naija as a valuable resource for future computational linguistic research.

The Linear State-Space Layer (LSSL) is a model family that combines the strengths of sequential modeling paradigms such as recurrence, convolution, and differential equations. For example, they generalize convolutions to continuous-time, explain common RNN heuristics, and share features of NDEs such as time-scale adaptation. Although naive LSSLs struggle with modeling long dependencies, we introduce a class of LSSL (SLLSSL), which overcomes these limitations by utilizing a trainable set of structured matrices that endow it with long range memory.

We present ongoing research on using data cartography techniques to better understand the training dynamics of neural sequence-to-sequence models for Machine Translation (MT). This is particularly needed to inform the design of effective training algorithms for MT in low-resource settings. Current results show that the difficulty of training samples depends on the current training phase and raises questions for future work, i.e. what are the properties of these samples and how can we modify the training process at each phase.

In recent years there have been great interests in addressing the low resourcefulness of African languages and providing baseline models for different Natural Language Processing tasks. Several initiatives on the continent use the Bible as a data source to provide proof of concept for some NLP tasks. In this work, we present the Lingala Speech Translation (LiSTra) dataset, release a full pipeline for the construction of such dataset in other languages, and report baselines using both the traditional cascade approach (Automatic Speech Recognition -> Machine Translation) and a revolutionary transformer-based End-2-End architecture with customized interactive attention that allows information sharing between the recognition decoder and the translation decoder.

The State Capture Commission of Inquiry has piqued the interest of a large number of viewers over the past few years. This research aims to help ease the burden of a user sifting through hundred of videos for a topic of interest through the proposal of a topic extraction based recommender system. With a large textual dataset at disposal, the research investigates the use of topic extraction through the combination of approaches such as Non-Negative Matrix Factorization and Latent Dirichlet Allocation for the construction of an algorithm that will be used for keyword based topic extraction. The result of this will allow the most relevant State Capture video related to a topic being presented to the user.

Designing robots that can learn multiple tasks is a difficult problem because learning from scratch for each task is prohibitively slow, leveraging experience from previous tasks might be difficult, and there might not exist a suitable reward function for RL. Thus some desirable properties are the ability to learn from demonstrations through imitation learning, which requires no reward function and can lead to exponential decreases in sample complexity, and to leverage past experience through meta-learning, which improves few-shot learning. Such robots would also benefit from leveraging hierarchy to re-use simple learned skills for more complex tasks, but learning hierarchical policies has historically proven difficult. In this work, we move closer towards satisfying these properties by formulating a hierarchical imitation learning problem which we tackle through meta-learning to leverage experience from multiple tasks. We present results on a linear point mass environment as well as a challenging simulated kitchen environment with a 7-DoF robotic arm.